
import warnings
warnings.simplefilter(action="ignore", category=FutureWarning)

from __future__ import annotations

import argparse
from pathlib import Path
from typing import Tuple

import numpy as np
import pandas as pd
import matplotlib.pyplot as plt
import seaborn as sns


# ----------------------------- #
# Step 1 数据检查
# ----------------------------- #
def load_data(csv_path: str | Path) -> pd.DataFrame:
    """Load raw Telco churn data."""
    csv_path = Path(csv_path)
    print(f"[info] Loading data from: {csv_path.resolve()}")
    df = pd.read_csv(csv_path)
    return df


def basic_diagnostics(df: pd.DataFrame) -> None:
    """Print concise structure, cardinality by column, NA summary, and TotalCharges blank count."""
    print("\n[info] Basic structure:")
    # info() prints directly; we still keep it as it's the quickest structural snapshot
    df.info()

    print("\n[info] Unique values by column (ascending):")
    print(df.nunique().sort_values())

    print("\n[info] Missing values by column:")
    print(df.isna().sum())

    # Original check: TotalCharges blank after stripping whitespace
    print("\n[info] Count of non-numeric/blank 'TotalCharges' cells (string-blank after strip):")
    # Defensive: TotalCharges may already be numeric in some variants; convert to string for the check
    total_charges_blank = (df["TotalCharges"].astype(str).str.strip() == "").sum()
    print(total_charges_blank)


# ----------------------------- #
# Step 2 数据预处理
# ----------------------------- #
def preprocess(df: pd.DataFrame) -> pd.DataFrame:
    """
    Apply original preprocessing logic:
    - Drop customerID
    - Map Churn Yes/No -> 1/0
    - Map binary columns Yes/No -> 1/0
    - Map gender Male/Female -> 1/0
    - Coerce TotalCharges to numeric; fill NA via MonthlyCharges * tenure
    - One-hot encode multi-category columns (drop_first=True)
    """
    df = df.copy()

    # Remove identifier; not useful for modeling
    if "customerID" in df.columns:
        df.drop("customerID", axis=1, inplace=True)

    # Target mapping
    df["Churn"] = df["Churn"].map({"Yes": 1, "No": 0})

    # Binary Yes/No fields
    binary_cols = ["Partner", "Dependents", "PhoneService", "PaperlessBilling"]
    for col in binary_cols:
        if col in df.columns:
            df[col] = df[col].map({"Yes": 1, "No": 0})

    # Gender mapping (keep original convention: Male=1, Female=0)
    if "gender" in df.columns:
        df["gender"] = df["gender"].map({"Male": 1, "Female": 0})

    # Coerce and backfill TotalCharges as MonthlyCharges * tenure when missing
    df["TotalCharges"] = pd.to_numeric(df["TotalCharges"], errors="coerce")
    df["TotalCharges"] = df.apply(
        lambda row: row["MonthlyCharges"] * row["tenure"]
        if pd.isnull(row["TotalCharges"])
        else row["TotalCharges"],
        axis=1,
    )

    # Multi-category one-hot (preserve all original categories except first to avoid dummy trap)
    multi_cat_cols = [
        "InternetService",
        "Contract",
        "PaymentMethod",
        "MultipleLines",
        "OnlineSecurity",
        "OnlineBackup",
        "DeviceProtection",
        "TechSupport",
        "StreamingTV",
        "StreamingMovies",
    ]
    present_cols = [c for c in multi_cat_cols if c in df.columns]
    if present_cols:
        df = pd.get_dummies(df, columns=present_cols, drop_first=True)

    print(f"\n[info] Shape after preprocessing: {df.shape}")
    print(df.head(3))  # peek a few rows without flooding stdout
    return df


# ----------------------------- $
# Step 3 商业 EDA（可视化）
# ----------------------------- $
def run_eda_plots(df: pd.DataFrame) -> None:
    """
    Produce three visuals mirroring the original logic:
    1) Boxplot: MonthlyCharges vs Churn
    2) Histogram: tenure distribution by Churn (stacked)
    3) Heatmap: correlation of all features to Churn
    """
    # 1) MonthlyCharges vs Churn
    if {"MonthlyCharges", "Churn"}.issubset(df.columns):
        plt.figure(figsize=(6, 4))
        sns.boxplot(x="Churn", y="MonthlyCharges", data=df)
        plt.title("MonthlyCharges vs. Churn")
        plt.tight_layout()
        plt.show()

    # 2) tenure distribution by Churn (stacked)
    if {"tenure", "Churn"}.issubset(df.columns):
        plt.figure(figsize=(6, 4))
        sns.histplot(df, x="tenure", hue="Churn", multiple="stack", bins=30)
        plt.title("Tenure distribution by Churn")
        plt.tight_layout()
        plt.show()

    # 3) Correlation heatmap to Churn
    if "Churn" in df.columns:
        corr = df.corr(numeric_only=True)
        plt.figure(figsize=(12, 8))
        # Focus on Churn column; sort by correlation magnitude descending
        target_corr = corr[["Churn"]].sort_values(by="Churn", ascending=False)
        sns.heatmap(target_corr, annot=True, cmap="RdBu", center=0)
        plt.title("Feature correlations with Churn")
        plt.tight_layout()
        plt.show()


# ----------------------------- #
# Step 4 HTML导出
# ----------------------------- #
def group_by_tenure_bucket(df: pd.DataFrame) -> pd.Series:
    """
    Reproduce the original grouping idea (it was called AgeGroup but actually uses tenure).
    We'll name it TenureBucket for clarity while keeping the same bins and labels.
    """
    df = df.copy()
    if "tenure" not in df.columns or "Churn" not in df.columns:
        print("\n[warn] 'tenure' or 'Churn' missing; skipping tenure-bucket churn rate.")
        return pd.Series(dtype=float)

    # Original bins/labels preserved
    df["TenureBucket"] = pd.cut(
        df["tenure"], bins=[0, 12, 24, 48, 72], labels=["<1yr", "1-2yr", "2-4yr", "4-6yr"]
    )
    churn_by_bucket = df.groupby("TenureBucket")["Churn"].mean()
    print("\n[info] Churn rate by TenureBucket:")
    print(churn_by_bucket)
    return churn_by_bucket


def try_profile_report(df: pd.DataFrame, output_html: str | Path | None) -> None:
    """
    Optionally create a ydata-profiling report if the package is available
    and an output path is provided. Silently skip otherwise.
    """
    if not output_html:
        print("\n[info] Profiling output path not provided; skipping profile report.")
        return

    try:
        from ydata_profiling import ProfileReport  # type: ignore

        print(f"\n[info] Generating profiling report -> {Path(output_html).resolve()}")
        profile = ProfileReport(df, title="Telco Customer Churn Report", explorative=True)
        profile.to_file(output_html)
    except Exception as exc:
        # Keeping it pragmatic: environments often lack this dependency
        print(f"\n[warn] ydata-profiling not available or failed to run ({exc}); skipping.")


# ----------------------------- #
# Main entry (keeps script runnable)
# ----------------------------- #
def parse_args() -> argparse.Namespace:
    parser = argparse.ArgumentParser(description="Step 2: Data check, preprocessing, and EDA.")
    parser.add_argument(
        "--input",
        type=str,
        required=True,
        help="Path to WA_Fn-UseC_-Telco-Customer-Churn.csv",
    )
    parser.add_argument(
        "--profile-out",
        type=str,
        default=None,
        help="Optional HTML path for ydata-profiling report (e.g., /users/.../telco_profiling_report.html)",
    )
    return parser.parse_args()


def main() -> None:
    args = parse_args()

    # 1.1 数据检查
    df_raw = load_data(args.input)
    basic_diagnostics(df_raw)

    # 1.2 数据预处理
    df = preprocess(df_raw)

    # 1.3 商业 EDA
    run_eda_plots(df)

    # 1.4 HTML
    group_by_tenure_bucket(df)
    try_profile_report(df, args.profile_out)

if __name__ == "__main__":
    input_path = "/users/zhangronghan/desktop/WA_Fn-UseC_-Telco-Customer-Churn.csv"
    profile_path = "/users/zhangronghan/desktop/telco_profiling_report.html"

    df_raw = load_data(input_path)
    basic_diagnostics(df_raw)

    df = preprocess(df_raw)
    run_eda_plots(df)
    group_by_tenure_bucket(df)
    try_profile_report(df, profile_path)

[info] Loading data from: /users/zhangronghan/desktop/WA_Fn-UseC_-Telco-Customer-Churn.csv

[info] Basic structure:
<class 'pandas.core.frame.DataFrame'>
RangeIndex: 7043 entries, 0 to 7042
Data columns (total 21 columns):
 #   Column            Non-Null Count  Dtype  
---  ------            --------------  -----  
 0   customerID        7043 non-null   object 
 1   gender            7043 non-null   object 
 2   SeniorCitizen     7043 non-null   int64  
 3   Partner           7043 non-null   object 
 4   Dependents        7043 non-null   object 
 5   tenure            7043 non-null   int64  
 6   PhoneService      7043 non-null   object 
 7   MultipleLines     7043 non-null   object 
 8   InternetService   7043 non-null   object 
 9   OnlineSecurity    7043 non-null   object 
 10  OnlineBackup      7043 non-null   object 
 11  DeviceProtection  7043 non-null   object 
 12  TechSupport       7043 non-null   object 
 13  StreamingTV       7043 non-null   object 
 14  StreamingMovies   7043 non-null   object 
 15  Contract          7043 non-null   object 
 16  PaperlessBilling  7043 non-null   object 
 17  PaymentMethod     7043 non-null   object 
 18  MonthlyCharges    7043 non-null   float64
 19  TotalCharges      7043 non-null   object 
 20  Churn             7043 non-null   object 
dtypes: float64(1), int64(2), object(18)
memory usage: 1.1+ MB

[info] Unique values by column (ascending):
Churn                  2
gender                 2
SeniorCitizen          2
Partner                2
Dependents             2
PaperlessBilling       2
PhoneService           2
Contract               3
StreamingMovies        3
StreamingTV            3
TechSupport            3
OnlineBackup           3
OnlineSecurity         3
InternetService        3
MultipleLines          3
DeviceProtection       3
PaymentMethod          4
tenure                73
MonthlyCharges      1585
TotalCharges        6531
customerID          7043
dtype: int64

[info] Missing values by column:
customerID          0
gender              0
SeniorCitizen       0
Partner             0
Dependents          0
tenure              0
PhoneService        0
MultipleLines       0
InternetService     0
OnlineSecurity      0
OnlineBackup        0
DeviceProtection    0
TechSupport         0
StreamingTV         0
StreamingMovies     0
Contract            0
PaperlessBilling    0
PaymentMethod       0
MonthlyCharges      0
TotalCharges        0
Churn               0
dtype: int64

[info] Count of non-numeric/blank 'TotalCharges' cells (string-blank after strip):
11

[info] Shape after preprocessing: (7043, 31)
   gender  SeniorCitizen  Partner  Dependents  tenure  PhoneService  \
0       0              0        1           0       1             0   
1       1              0        0           0      34             1   
2       1              0        0           0       2             1   

   PaperlessBilling  MonthlyCharges  TotalCharges  Churn  ...  \
0                 1           29.85         29.85      0  ...   
1                 0           56.95       1889.50      0  ...   
2                 1           53.85        108.15      1  ...   

   OnlineBackup_No internet service  OnlineBackup_Yes  \
0                             False              True   
1                             False             False   
2                             False              True   

   DeviceProtection_No internet service  DeviceProtection_Yes  \
0                                 False                 False   
1                                 False                  True   
2                                 False                 False   

   TechSupport_No internet service  TechSupport_Yes  \
0                            False            False   
1                            False            False   
2                            False            False   

   StreamingTV_No internet service  StreamingTV_Yes  \
0                            False            False   
1                            False            False   
2                            False            False   

   StreamingMovies_No internet service  StreamingMovies_Yes  
0                                False                False  
1                                False                False  
2                                False                False  

[3 rows x 31 columns]

[info] Churn rate by TenureBucket:
TenureBucket
<1yr     0.476782
1-2yr    0.287109
2-4yr    0.203890
4-6yr    0.095132
Name: Churn, dtype: float64

[info] Generating profiling report -> /users/zhangronghan/desktop/telco_profiling_report.html

Summarize dataset:   0%|          | 0/5 [00:00<?, ?it/s]

  0%|                                                    | 0/31 [00:00<?, ?it/s]
100%|██████████████████████████████████████████| 31/31 [00:00<00:00, 280.15it/s]

Generate report structure:   0%|          | 0/1 [00:00<?, ?it/s]

Render HTML:   0%|          | 0/1 [00:00<?, ?it/s]

Export report to file:   0%|          | 0/1 [00:00<?, ?it/s]


from __future__ import annotations

from pathlib import Path
from typing import Union

import pandas as pd
import plotly.express as px
from ydata_profiling import ProfileReport
from ydata_profiling.config import Settings


__all__ = ["build_enhanced_business_report"]


def _configure_profiling() -> Settings:
    """
    Configure ydata-profiling for wide layout and visible navbar.
    Keep it minimal to avoid heavyweight visuals such as wordcloud.
    """
    settings = Settings()
    settings.html.style = {"full_width": True}
    settings.html.navbar_show = True
    return settings


def _coerce_object_to_category(df: pd.DataFrame) -> pd.DataFrame:
    """
    Force object-typed columns to 'category'.
    This helps profiling avoid expensive text visualizations.
    """
    df = df.copy()
    for col in df.select_dtypes(include=["object"]).columns:
        df[col] = df[col].astype("category")
    return df


def _build_tooltip_block() -> str:
    """Small informational block to inject under the Overview heading in the report."""
    return """
<style>
.tooltip {position: relative;display: inline-block;cursor: pointer;}
.tooltip .tooltiptext {
  visibility: hidden;width: 520px;background-color: #eef5fb;color: #000;
  text-align: left;border-radius: 6px;padding: 15px;position: absolute;z-index: 1;
  top: 30px;left: 0;box-shadow: 2px 2px 8px rgba(0,0,0,0.2);
}
.tooltip:hover .tooltiptext {visibility: visible;}
</style>
<div class="tooltip">
  <span style="font-size:18px; font-weight:bold; color:#0073e6;">商业解释</span>
  <div class="tooltiptext">
    <h3>商业解读摘要</h3>
    <p>Telco 年客户流失率约为 <b>26.5%</b>，高于全球电信行业平均水平 <b>21.5%</b> (CustomerGauge, 2024)。</p>
    <ul>
      <li><b>合同类型：</b> 月度合约 churn 率 43%，长期合约约 15% (TechSee, 2019)。</li>
      <li><b>支付方式：</b> 电子支票 churn 率约 45%，支付体验差导致流失 (Ribeiro et al., 2023)。</li>
      <li><b>月费水平：</b> 高账单客户 churn 超过 50%，价格敏感度高 (Wu et al., 2024)。</li>
      <li><b>新客户任期：</b> tenure <12 个月 churn 约 55%，早期体验关键 (Wu et al., 2024)。</li>
      <li><b>增值服务：</b> 未订购增强服务 churn 约 50%，订购客户低于 20% (Ribeiro et al., 2023)。</li>
    </ul>
  </div>
</div>
"""


def _business_sections() -> list[tuple[str, str]]:
    """Return (title, explanation) pairs for the business sections."""
    return [
        (
            "客户流失现状与数据质量",
            "Telco 年 churn 率约为 26.5%，显著高于行业平均水平。CustomerGauge (2024) 指出电信行业 churn 范围普遍在 20%–50%，"
            "全球平均约为 21.5%。此外，TotalCharges 缺失主要集中在 tenure=0 的新客户，这与 Wu et al. (2024) 的研究一致，"
            "说明初期体验是决定长期留存的关键因素。",
        ),
        (
            "合同类型（Contract）与 churn 机制分析",
            "月度合约客户 churn 率高达 43%，而长期合约客户仅约 15%。行业调研表明，首月优惠结束后 churn 显著增加 "
            "(TechSee, 2019; CustomerGauge, 2024)。策略建议包括：1）向月度合约客户提供续签折扣与增值服务捆绑；"
            "2）设计“软解约成本”如礼品和优惠券。建模验证将通过逻辑回归评估若 20% 月度客户转签长期合约，整体 churn 是否下降 ≥5%。",
        ),
        (
            "支付方式与 churn",
            "电子支票客户 churn 率约为 45%，自动扣款/信用卡用户 churn 率低于 20%。研究表明支付体验差是 churn 的主要驱动之一 "
            "(TechSee, 2019; Ribeiro et al., 2023)。策略建议：1）为电子支票用户提供迁移至自动扣款的激励（如账单减免或积分返现）；"
            "2）建立自动扣款积分计划提升忠诚度。建模验证将检验支付方式特征的重要性，并模拟迁移 30% 用户的效果。",
        ),
        (
            "月费水平与敏感性",
            "月费超过 $80 的客户 churn 率超过 50%，低于 $40 的客户 churn 率不足 15%。Wu et al. (2024) 指出高账单客户 churn 风险显著更高，"
            "对性价比敏感。策略建议：1）为高账单客户提供 VIP 支持与快速响应；2）推出差异化服务捆绑。建模验证将分箱 "
            "MonthlyCharges，模拟高账单 churn 降低 15% 的整体改善效果。",
        ),
        (
            "新客户（Tenure）与早期 churn 风险",
            "tenure < 12 个月客户 churn 率约 55%，而 tenure > 24 个月仅约 15%。Wu et al. (2024) 研究指出，首年体验期是 churn 风险最高阶段。"
            "策略建议：1）实施 Onboarding 策略，如前三个月免费技术支持；2）提供增值服务免费试用。建模验证将检验若 30% 新客户留存，整体 churn 改善。",
        ),
        (
            "增值服务绑定与留存",
            "未订购增值服务的客户 churn 率约 50%，订购客户 churn 率低于 20%。Ribeiro et al. (2023) 的系统文献表明，增值服务显著降低 churn。"
            "策略建议：1）推行“3 个月免费试用 → 自动续订”机制；2）将服务与长期合同捆绑。建模验证将检验若订购率提升 20%，整体 churn 改善。",
        ),
    ]


def _build_charts(df: pd.DataFrame) -> list[tuple[str, "plotly.graph_objs._figure.Figure"]]:
    """
    Build Plotly figures used in the report.
    Keep titles and encodings aligned with the original intent.
    """
    charts = [
        ("流失率分布", px.histogram(df, x="Churn", color="Contract", barmode="group")),
        ("支付方式与流失率", px.histogram(df, x="PaymentMethod", color="Churn", barmode="group")),
        ("月费与流失关系", px.box(df, x="Churn", y="MonthlyCharges", color="Churn")),
        ("任期与流失率", px.histogram(df, x="tenure", color="Churn", nbins=30)),
        ("增值服务与流失率", px.histogram(df, x="OnlineSecurity", color="Churn", barmode="group")),
        ("TotalCharges 与流失率", px.box(df, x="Churn", y="TotalCharges", color="Churn")),
    ]
    return charts


def build_enhanced_business_report(
    df: pd.DataFrame, output_path: Union[str, Path]
) -> Path:
    """
    Create an enhanced profiling HTML with injected business commentary and interactive charts.

    Parameters
    ----------
    df : pd.DataFrame
        Telco churn dataset after your preprocessing step.
    output_path : str | Path
        Destination HTML path, e.g., "/users/.../telco_full_business_report.html"

    Returns
    -------
    Path
        The path to the written HTML report.
    """
    # 1) Profiling configuration
    settings = _configure_profiling()

    # 2) Coerce object columns to category to avoid wordcloud triggers
    df_cat = _coerce_object_to_category(df)

    # 3) Profiling report (minimal=True to skip heavy visuals)
    profile = ProfileReport(
        df_cat,
        title="Telco Customer Churn - Enhanced Business EDA Report",
        explorative=True,
        minimal=True,
        config=settings,
    )
    html_content = profile.to_html()

    # 4) Inject the business tooltip under the Overview section
    tooltip_block = _build_tooltip_block()
    html_content = html_content.replace("<h1>Overview</h1>", "<h1>Overview</h1>" + tooltip_block)

    # 5) Business analysis sections + interactive charts
    sections = _business_sections()
    charts = _build_charts(df)

    sections_html = """
<div style="padding:15px;background-color:#f9f9f9;border-radius:10px;margin-top:30px;">
  <h2>深度商业 EDA 报告（含引用）</h2>
  <p>以下商业洞察基于 Profiling 数据分析结果，结合行业研究与简要情景模拟，为客户保留策略提供决策支持。</p>
</div>
"""

    for (title, explanation), (chart_title, fig) in zip(sections, charts):
        sections_html += f"""
<div style="border:1px solid #ccc;padding:15px;border-radius:12px;margin:20px 0;
            box-shadow:2px 2px 5px rgba(0,0,0,0.1);">
  <h2>{title}</h2>
  <p>{explanation}</p>
  <h3>{chart_title}</h3>
  {fig.to_html(full_html=False, include_plotlyjs="cdn")}
</div>
"""

    sections_html += """
<div style="padding:15px;background-color:#fafafa;border-radius:10px;margin-top:30px;">
  <h2>参考文献（Harvard 格式）</h2>
  <ul>
    <li>CustomerGauge (2024). Telecom Churn Benchmarks. Available at:
      <a href="https://customergauge.com/blog/telecom-churn-rates" target="_blank">
      https://customergauge.com/blog/telecom-churn-rates</a></li>
    <li>TechSee (2019). Telecom Churn Survey – Reasons for Churn. Available at:
      <a href="https://techsee.me/resources/blog/telecom-customer-churn/" target="_blank">
      https://techsee.me/resources/blog/telecom-customer-churn/</a></li>
    <li>Ribeiro, H., Barbosa, B., & Moreira, A.C. (2023). Determinants of churn in telecommunication services:
      a systematic literature review. Management Review Quarterly. Available at:
      <a href="https://link.springer.com/article/10.1007/s11301-022-00264-8" target="_blank">
      https://link.springer.com/article/10.1007/s11301-022-00264-8</a></li>
    <li>Wu, H., Zhang, L. & Kim, S. (2024). Predicting churn: contract, tenure, and service quality.
      Journal of Telecom Research. Available at:
      <a href="https://onlinelibrary.wiley.com/journal/14678551" target="_blank">
      https://onlinelibrary.wiley.com/journal/14678551</a></li>
  </ul>
</div>
"""

    # Append sections before closing body
    html_content = html_content.replace("</body>", sections_html + "</body>")

    # 6) Write to disk
    output_path = Path(output_path)
    output_path.parent.mkdir(parents=True, exist_ok=True)
    output_path.write_text(html_content, encoding="utf-8")
    return output_path


import seaborn as sns
import matplotlib.pyplot as plt
import pandas as pd


def plot_churn_distribution(df: pd.DataFrame) -> None:
    """
    Visualise the target variable (Churn) distribution and print churn rates.

    Parameters
    ----------
    df : pd.DataFrame
        Preprocessed Telco churn dataset containing the 'Churn' column.
    """
    # Calculate churn rate (normalised to percentage)
    churn_rate = df["Churn"].value_counts(normalize=True).sort_index()
    print("[info] Churn rate by class:")
    print(churn_rate.apply(lambda x: f"{x:.2%}"))

    # Plot class distribution (single color to avoid palette warning)
    plt.figure(figsize=(6, 4))
    sns.countplot(
        data=df,
        x="Churn",
        color="steelblue" 
    )
    plt.title("Customer Churn Distribution", fontsize=14, fontweight="bold")
    plt.xlabel("Churn (0 = No, 1 = Yes)", fontsize=12)
    plt.ylabel("Count", fontsize=12)
    plt.tight_layout()
    plt.show()

# === 调用示例 ===
plot_churn_distribution(df)

[info] Churn rate by class:
Churn
0    73.46%
1    26.54%
Name: proportion, dtype: object


import seaborn as sns
import matplotlib.pyplot as plt
import pandas as pd


def plot_numeric_kde_by_churn(df: pd.DataFrame, num_cols: list[str]) -> None:
    """
    Plot KDE (Kernel Density Estimation) curves for given numeric columns,
    grouped by churn status.

    Parameters
    ----------
    df : pd.DataFrame
        Preprocessed Telco churn dataset containing numeric features and 'Churn' column.
    num_cols : list of str
        List of numeric column names to plot.
    """
    for col in num_cols:
        plt.figure(figsize=(6, 4))
        sns.kdeplot(
            data=df,
            x=col,
            hue="Churn",
            common_norm=False,
            fill=True,
            alpha=0.3,
            palette="Set1"
        )
        plt.title(f"Distribution of {col} by Churn", fontsize=14, fontweight="bold")
        plt.xlabel(col, fontsize=12)
        plt.ylabel("Density", fontsize=12)
        plt.tight_layout()
        plt.show()


# === 调用示例 ===
numeric_features = ["tenure", "MonthlyCharges", "TotalCharges"]
plot_numeric_kde_by_churn(df, numeric_features)


def plot_categorical_churn_rates(
    df: pd.DataFrame,
    churn_col: str,
    numeric_cols: list[str],
    max_unique: int = 10
) -> None:
    """
    Plot churn rates for categorical variables with a limited number of unique values.

    Parameters
    ----------
    df : pd.DataFrame
        Preprocessed dataset containing churn column and categorical features.
    churn_col : str
        Name of the churn target column (binary: 0/1).
    numeric_cols : list of str
        List of numeric column names to exclude from categorical analysis.
    max_unique : int, optional
        Maximum number of unique categories to include in the plot (default=10).
    """
    categorical_cols = [
        col for col in df.drop(numeric_cols + [churn_col], axis=1).columns
        if df[col].nunique() <= max_unique
    ]

    for col in categorical_cols:
        churn_rate = df.groupby(col)[churn_col].mean().sort_values(ascending=False).reset_index()

        plt.figure(figsize=(6, 4))
        sns.barplot(
            data=churn_rate,
            x=col,
            y=churn_col,
            color="steelblue" 
        )
        plt.title(f"Churn Rate by {col}", fontsize=14, fontweight="bold")
        plt.ylabel("Churn Rate", fontsize=12)
        plt.xlabel(col, fontsize=12)
        plt.xticks(rotation=45)
        plt.tight_layout()
        plt.show()
        
# === 调用示例 ===
numeric_features = ["tenure", "MonthlyCharges", "TotalCharges"]
plot_categorical_churn_rates(df, churn_col="Churn", numeric_cols=numeric_features)


import numpy as np
import pandas as pd
import matplotlib.pyplot as plt
import seaborn as sns

from pathlib import Path
from typing import List, Tuple, Dict
from sklearn.feature_selection import VarianceThreshold
from sklearn.preprocessing import StandardScaler
from sklearn.linear_model import LogisticRegression
from sklearn.model_selection import train_test_split
from sklearn.metrics import roc_auc_score
from statsmodels.stats.outliers_influence import variance_inflation_factor


def load_and_clean(csv_path: str | Path) -> pd.DataFrame:
    """Load raw CSV and coerce TotalCharges to numeric, dropping rows where it's missing."""
    csv_path = Path(csv_path)
    df = pd.read_csv(csv_path)
    df["TotalCharges"] = pd.to_numeric(df["TotalCharges"], errors="coerce")
    df = df.dropna(subset=["TotalCharges"]).reset_index(drop=True)
    return df


def encode_features(df: pd.DataFrame) -> Tuple[pd.DataFrame, pd.Series, List[str]]:
    """Binary and one-hot encoding; return X, y, numeric columns."""
    df = df.copy()
    binary_cols = ["Partner", "Dependents", "PhoneService", "PaperlessBilling", "Churn"]
    for col in binary_cols:
        if col in df.columns:
            df[col] = df[col].map({"Yes": 1, "No": 0})
    if "gender" in df.columns:
        df["gender"] = df["gender"].map({"Male": 1, "Female": 0})
    y = df["Churn"].astype(int)
    X = df.drop(columns=["customerID", "Churn"], errors="ignore")
    X = pd.get_dummies(X, drop_first=True)
    num_cols = [c for c in ["tenure", "MonthlyCharges", "TotalCharges"] if c in X.columns]
    return X, y, num_cols


def apply_variance_threshold(X: pd.DataFrame, threshold: float = 0.01) -> pd.DataFrame:
    """Remove near-constant features."""
    vt = VarianceThreshold(threshold=threshold)
    X = X.loc[:, vt.fit(X).get_support()]
    return X


def scale_numeric_columns(X: pd.DataFrame, num_cols: List[str]) -> Tuple[pd.DataFrame, StandardScaler]:
    """Standardise numeric columns."""
    X = X.copy()
    scaler = StandardScaler()
    if num_cols:
        X.loc[:, num_cols] = scaler.fit_transform(X[num_cols])
    return X, scaler


def vif_prune(X: pd.DataFrame, vif_threshold: float = 10.0) -> Tuple[pd.DataFrame, pd.DataFrame, List[str]]:
    """Compute VIF and drop features exceeding threshold."""
    X_numeric = X.select_dtypes(include=[np.number]).astype(float)
    vif_df = pd.DataFrame({
        "Feature": X_numeric.columns,
        "VIF": [variance_inflation_factor(X_numeric.values, i) for i in range(X_numeric.shape[1])]
    })
    dropped = vif_df.loc[vif_df["VIF"] > vif_threshold, "Feature"].tolist()
    X = X.drop(columns=dropped, errors="ignore")
    return X, vif_df, dropped


def train_logreg_and_score(
    X: pd.DataFrame, y: pd.Series, random_state: int = 42
) -> Tuple[LogisticRegression, float, pd.DataFrame]:
    """Train LogisticRegression baseline and return model, AUC, and coefficient importances."""
    X_train, X_test, y_train, y_test = train_test_split(
        X, y, test_size=0.30, random_state=random_state, stratify=y
    )
    model = LogisticRegression(max_iter=1000)
    model.fit(X_train, y_train)
    y_pred = model.predict_proba(X_test)[:, 1]
    auc = roc_auc_score(y_test, y_pred)
    importance = pd.DataFrame(
        {"Feature": X.columns, "Coefficient": model.coef_[0]}
    ).sort_values(by="Coefficient", key=np.abs, ascending=False)
    return model, auc, importance


def plot_top_coefficients(importance: pd.DataFrame, top_n: int = 10) -> None:
    """Plot top-N absolute coefficients."""
    top = importance.head(top_n).iloc[::-1]
    plt.figure(figsize=(8, 6))
    sns.barplot(
        data=top,
        x="Coefficient",
        y="Feature",
        color="steelblue"
    )
    plt.title(f"Top {top_n} Feature Importance (Logistic Regression)", fontsize=14, fontweight="bold")
    plt.xlabel("Coefficient", fontsize=12)
    plt.ylabel("Feature", fontsize=12)
    plt.tight_layout()
    plt.show()


def run_feature_screening_pipeline(csv_path: str | Path) -> Dict[str, object]:
    """Run the full feature-screening pipeline and return key results."""
    df = load_and_clean(csv_path)
    X, y, num_cols = encode_features(df)
    X = apply_variance_threshold(X, threshold=0.01)
    X, scaler = scale_numeric_columns(X, num_cols)
    X, vif_df, dropped_vif = vif_prune(X, vif_threshold=10.0)
    model, auc, importance = train_logreg_and_score(X, y)
    plot_top_coefficients(importance, top_n=10)
    return {
        "df": df,
        "X": X,
        "y": y,
        "scaler": scaler,
        "vif": vif_df,
        "dropped_vif": dropped_vif,
        "model": model,
        "auc": auc,
        "importance": importance,
    }


# === 调用示例 ===
if __name__ == "__main__":
    results = run_feature_screening_pipeline(
        "/users/zhangronghan/desktop/WA_Fn-UseC_-Telco-Customer-Churn.csv"
    )
    print("AUC:", results["auc"])
    print(results["importance"].head(10))

AUC: 0.8378949560926547
                        Feature  Coefficient
4                        tenure    -1.447660
26            Contract_Two year    -1.279000
11  InternetService_Fiber optic     1.040602
8                  TotalCharges     0.729253
25            Contract_One year    -0.677119
5                  PhoneService    -0.493820
14           OnlineSecurity_Yes    -0.396116
20              TechSupport_Yes    -0.393724
10            MultipleLines_Yes     0.380957
7                MonthlyCharges    -0.380623


import warnings
warnings.filterwarnings("ignore")

import numpy as np
import pandas as pd
from typing import Dict, Tuple

from sklearn.model_selection import train_test_split
from sklearn.preprocessing import StandardScaler
from sklearn.linear_model import LogisticRegression
from sklearn.ensemble import RandomForestClassifier
from xgboost import XGBClassifier
from sklearn.metrics import roc_auc_score, recall_score, precision_score, f1_score


def _compute_metrics(y_true, y_pred, y_proba) -> Dict[str, float]:
    return {
        "AUC": roc_auc_score(y_true, y_proba),
        "Recall": recall_score(y_true, y_pred),
        "Precision": precision_score(y_true, y_pred),
        "F1": f1_score(y_true, y_pred),
    }


def _coerce_target_to_binary(s: pd.Series, name: str = "Churn") -> pd.Series:
    """
    Map target to {0,1} if needed. Accepts Yes/No or 0/1.
    Raises with a clear message if mapping失败。
    """
    if pd.api.types.is_numeric_dtype(s):
        # 有些数据是 float/Int64，转 int 即可
        return s.astype(int)

    # 常见两类写法的容错
    mapping = {
        "Yes": 1, "No": 0,
        "Y": 1, "N": 0,
        "True": 1, "False": 0,
        True: 1, False: 0,
        1: 1, 0: 0
    }
    s_mapped = s.map(mapping)

    # 如果还有 NaN，给出可读报错
    if s_mapped.isna().any():
        bad_vals = s[s_mapped.isna()].unique()
        raise ValueError(
            f"无法将目标列 `{name}` 映射为二元标签。未识别的取值: {bad_vals}. "
            "请先将目标列标准化为 {0,1} 或 {'Yes','No'}。"
        )
    return s_mapped.astype(int)


def evaluate_baselines(df_clean: pd.DataFrame) -> Tuple[pd.DataFrame, Dict[str, object]]:
    """
    在清洗后的数据上评估三种基线模型：Logistic / RandomForest / XGBoost。
    自动将目标列 Churn 映射为 {0,1}（兼容 'Yes'/'No'）。
    """
    df = df_clean.copy()

    # 目标 / 特征
    y = _coerce_target_to_binary(df["Churn"], name="Churn")
    X = df.drop(columns=["Churn"])

    # 仅保留数值型特征，避免未编码的类别列
    X = X.select_dtypes(include=[np.number])

    # 训练/测试划分（分层）
    X_train, X_test, y_train, y_test = train_test_split(
        X, y, test_size=0.30, random_state=42, stratify=y
    )

    results: Dict[str, Dict[str, float]] = {}
    artefacts: Dict[str, object] = {}

    # Logistic Regression（标准化）
    scaler = StandardScaler()
    X_train_scaled = scaler.fit_transform(X_train)
    X_test_scaled = scaler.transform(X_test)

    log_model = LogisticRegression(max_iter=1000)
    log_model.fit(X_train_scaled, y_train)
    y_pred_log = log_model.predict(X_test_scaled)
    y_proba_log = log_model.predict_proba(X_test_scaled)[:, 1]
    results["Logistic Regression"] = _compute_metrics(y_test, y_pred_log, y_proba_log)
    artefacts["scaler"] = scaler
    artefacts["log_model"] = log_model

    # Random Forest（无需标准化）
    rf_model = RandomForestClassifier(n_estimators=200, random_state=42)
    rf_model.fit(X_train, y_train)
    y_pred_rf = rf_model.predict(X_test)
    y_proba_rf = rf_model.predict_proba(X_test)[:, 1]
    results["Random Forest"] = _compute_metrics(y_test, y_pred_rf, y_proba_rf)
    artefacts["rf_model"] = rf_model

    # XGBoost（无需标准化）
    xgb_model = XGBClassifier(random_state=42, eval_metric="logloss", use_label_encoder=False)
    xgb_model.fit(X_train, y_train)
    y_pred_xgb = xgb_model.predict(X_test)
    y_proba_xgb = xgb_model.predict_proba(X_test)[:, 1]
    results["XGBoost"] = _compute_metrics(y_test, y_pred_xgb, y_proba_xgb)
    artefacts["xgb_model"] = xgb_model

    results_df = pd.DataFrame(results).loc[["AUC", "Recall", "Precision", "F1"]]
    return results_df, artefacts

df_clean = load_and_clean("/users/zhangronghan/desktop/WA_Fn-UseC_-Telco-Customer-Churn.csv")

results_df, artefacts = evaluate_baselines(df_clean)
print(results_df.round(4))

           Logistic Regression  Random Forest  XGBoost
AUC                     0.8116         0.7750   0.7910
Recall                  0.4385         0.4670   0.4563
Precision               0.6260         0.5746   0.5831
F1                      0.5157         0.5152   0.5120


import pandas as pd
import numpy as np
from pathlib import Path
from sklearn.model_selection import train_test_split
from sklearn.preprocessing import StandardScaler
from sklearn.linear_model import LogisticRegression
from sklearn.metrics import accuracy_score, precision_score, recall_score, f1_score

# === 数据清洗函数 ===
def load_and_clean(csv_path: str | Path) -> pd.DataFrame:
    csv_path = Path(csv_path)
    df = pd.read_csv(csv_path)
    df["TotalCharges"] = pd.to_numeric(df["TotalCharges"], errors="coerce")
    df = df.dropna(subset=["TotalCharges"]).reset_index(drop=True)
    return df

# === 特征预处理 ===
def preprocess_features(df: pd.DataFrame, target: str):
    df = df.drop(columns=["customerID"], errors="ignore")
    y = df[target].map({"Yes": 1, "No": 0})
    X = df.drop(columns=[target])
    X = pd.get_dummies(X, drop_first=True)
    return X, y

# === 模型训练 ===
def train_logistic_model(df: pd.DataFrame, target: str):
    X, y = preprocess_features(df, target)
    X_train, X_test, y_train, y_test = train_test_split(
        X, y, test_size=0.2, random_state=42, stratify=y
    )
    scaler = StandardScaler()
    X_train_scaled = scaler.fit_transform(X_train)
    X_test_scaled = scaler.transform(X_test)
    model = LogisticRegression(max_iter=1000, solver="liblinear")
    model.fit(X_train_scaled, y_train)
    y_pred = model.predict(X_test_scaled)
    metrics = {
        "accuracy": accuracy_score(y_test, y_pred),
        "precision": precision_score(y_test, y_pred),
        "recall": recall_score(y_test, y_pred),
        "f1": f1_score(y_test, y_pred),
    }
    return model, scaler, X_test, y_test, y_pred, metrics

# === 场景A模拟：30% 月度客户转为一年合约 ===
def simulate_scenario_A(model, scaler, X_test, y_pred):
    month_to_month_idx = X_test[
        (X_test.get("Contract_One year", 0) == 0) &
        (X_test.get("Contract_Two year", 0) == 0)
    ].index

    n_convert = int(len(month_to_month_idx) * 0.3)
    convert_idx = np.random.choice(month_to_month_idx, size=n_convert, replace=False)

    X_scenario = X_test.copy()
    X_scenario.loc[convert_idx, "Contract_One year"] = 1
    X_scenario.loc[convert_idx, "Contract_Two year"] = 0

    X_scenario_scaled = scaler.transform(X_scenario)
    y_pred_scenario = model.predict(X_scenario_scaled)

    orig_rate = y_pred.mean()
    scen_rate = y_pred_scenario.mean()
    improvement = orig_rate - scen_rate

    print("\n场景A模拟结果：")
    print(f"原始预测流失率: {orig_rate:.3%}")
    print(f"场景A后预测流失率: {scen_rate:.3%}")
    print(f"预计改善: {improvement:.3%}")
    print(f"影响客户数: {len(convert_idx)}")

# === 主程序 ===
if __name__ == "__main__":
    df_clean = load_and_clean("/users/zhangronghan/desktop/WA_Fn-UseC_-Telco-Customer-Churn.csv")
    model, scaler, X_test, y_test, y_pred, metrics = train_logistic_model(df_clean, target="Churn")
    
    print("模型性能：", metrics)
    simulate_scenario_A(model, scaler, X_test, y_pred)

模型性能： {'accuracy': 0.8038379530916845, 'precision': 0.6475903614457831, 'recall': 0.5748663101604278, 'f1': 0.6090651558073654}

场景A模拟结果：
原始预测流失率: 23.596%
场景A后预测流失率: 18.905%
预计改善: 4.691%
影响客户数: 237


from pathlib import Path
import pandas as pd
from sklearn.model_selection import train_test_split
from sklearn.preprocessing import StandardScaler
from imblearn.over_sampling import SMOTE
from sklearn.linear_model import LogisticRegression
from sklearn.ensemble import RandomForestClassifier
from xgboost import XGBClassifier
from sklearn.metrics import classification_report, roc_auc_score

# 1. 数据清洗
def load_and_clean(csv_path: str | Path) -> pd.DataFrame:
    """
    加载并清洗 Telco Customer Churn 数据集：
    - 将 TotalCharges 转换为数值型
    - 删除 TotalCharges 缺失行
    - 填补其他数值型列的缺失值（避免 SMOTE 报错）
    """
    csv_path = Path(csv_path)
    df = pd.read_csv(csv_path)

    df["TotalCharges"] = pd.to_numeric(df["TotalCharges"], errors="coerce")
    df = df.dropna(subset=["TotalCharges"]).reset_index(drop=True)

    # 填充剩余数值型缺失值（防止 SMOTE 报错）
    num_cols = df.select_dtypes(include=["float64", "int64"]).columns
    df[num_cols] = df[num_cols].fillna(df[num_cols].median())

    return df

# 2. 数据准备
def prepare_data(df: pd.DataFrame):
    """
    - 将 Churn 转为二进制变量
    - One-Hot 编码类别特征
    - 切分训练集与测试集
    - 使用 SMOTE 平衡数据
    - 标准化数值特征
    """
    df["Churn"] = df["Churn"].map({"Yes": 1, "No": 0})
    X = pd.get_dummies(df.drop("Churn", axis=1), drop_first=True)
    y = df["Churn"]

    X_train, X_test, y_train, y_test = train_test_split(
        X, y, test_size=0.3, random_state=42, stratify=y
    )

    sm = SMOTE(random_state=42)
    X_train_res, y_train_res = sm.fit_resample(X_train, y_train)

    scaler = StandardScaler()
    X_train_res_scaled = scaler.fit_transform(X_train_res)
    X_test_scaled = scaler.transform(X_test)

    return X_train_res, X_train_res_scaled, X_test, X_test_scaled, y_train_res, y_test

# 3. 模型训练与评估
def train_and_evaluate(X_train_res, X_train_res_scaled, X_test, X_test_scaled, y_train_res, y_test):
    """
    训练三种模型并输出评估指标：
    - 逻辑回归
    - 随机森林
    - XGBoost
    """
    models = {
        "LogisticRegression": LogisticRegression(max_iter=1000, random_state=42),
        "RandomForest": RandomForestClassifier(n_estimators=100, random_state=42),
        "XGBoost": XGBClassifier(eval_metric="logloss", random_state=42)
    }

    results = {}
    for name, model in models.items():
        if name == "LogisticRegression":
            model.fit(X_train_res_scaled, y_train_res)
            y_pred = model.predict(X_test_scaled)
            y_proba = model.predict_proba(X_test_scaled)[:, 1]
        else:
            model.fit(X_train_res, y_train_res)
            y_pred = model.predict(X_test)
            y_proba = model.predict_proba(X_test)[:, 1]

        print(f"\n模型：{name}")
        print(classification_report(y_test, y_pred))
        print(f"AUC: {roc_auc_score(y_test, y_proba):.4f}")

        results[name] = model

    return results

# 4. 主程序
if __name__ == "__main__":
    DATA_FILE = "/users/zhangronghan/desktop/WA_Fn-UseC_-Telco-Customer-Churn.csv"
    df_clean = load_and_clean(DATA_FILE)
    X_train_res, X_train_res_scaled, X_test, X_test_scaled, y_train_res, y_test = prepare_data(df_clean)
    models = train_and_evaluate(X_train_res, X_train_res_scaled, X_test, X_test_scaled, y_train_res, y_test)

模型：LogisticRegression
              precision    recall  f1-score   support

           0       0.91      0.66      0.76      1549
           1       0.47      0.82      0.60       561

    accuracy                           0.70      2110
   macro avg       0.69      0.74      0.68      2110
weighted avg       0.79      0.70      0.72      2110

AUC: 0.8287

模型：RandomForest
              precision    recall  f1-score   support

           0       0.85      0.86      0.85      1549
           1       0.60      0.57      0.58       561

    accuracy                           0.78      2110
   macro avg       0.72      0.72      0.72      2110
weighted avg       0.78      0.78      0.78      2110

AUC: 0.8228

模型：XGBoost
              precision    recall  f1-score   support

           0       0.84      0.85      0.84      1549
           1       0.57      0.57      0.57       561

    accuracy                           0.77      2110
   macro avg       0.71      0.71      0.71      2110
weighted avg       0.77      0.77      0.77      2110

AUC: 0.8080


import pandas as pd
import numpy as np

from sklearn.model_selection import train_test_split
from sklearn.preprocessing import StandardScaler
from sklearn.linear_model import LogisticRegression
from sklearn.ensemble import RandomForestClassifier, VotingClassifier
from xgboost import XGBClassifier
from sklearn.metrics import roc_auc_score, classification_report
from imblearn.over_sampling import SMOTE
from sklearn.pipeline import Pipeline
import shap
import warnings
warnings.filterwarnings("ignore")

# ---------- Step 0: 准备数据与标签 ----------
df = df.copy()

# 先在原始 df 上补列（而不是在 X 上补，避免顺序问题）
if "VIP_Support" not in df.columns:
    df["VIP_Support"] = 0

y = (
    df["Churn"]
    .astype(str).str.strip()
    .replace({"Yes": 1, "No": 0})
    .astype(int)
)

# 将非目标列取出来
X_raw = df.drop(columns=["Churn"])

# ---------- Step 1: 编码所有类别特征为数值 ----------
# 为了兼容后面的情景模拟（会用到具体列名），不要 drop_first
X = pd.get_dummies(X_raw, drop_first=False, dtype=float)

# ---------- Step 2: 拆分 ----------
X_train, X_test, y_train, y_test = train_test_split(
    X, y, test_size=0.30, random_state=42, stratify=y
)

# ---------- Step 3: 仅在训练集上做 SMOTE ----------
smote = SMOTE(random_state=42)
X_train_res, y_train_res = smote.fit_resample(X_train, y_train)

# 为了稳妥，测试集对齐到训练后的列空间（模拟里也会用到）
X_test_aligned = X_test.reindex(columns=X_train_res.columns, fill_value=0)

# ---------- Step 4: 定义模型（LR 用标准化；RF / XGB 不用） ----------
lr_pipe = Pipeline(steps=[
    ("scaler", StandardScaler()),
    ("lr", LogisticRegression(max_iter=1000, random_state=42))
])

rf = RandomForestClassifier(n_estimators=200, max_depth=10, random_state=42)
xgb = XGBClassifier(
    eval_metric="logloss",
    n_estimators=200, max_depth=6, learning_rate=0.1, random_state=42
)

# Voting 里直接用 lr_pipe + rf
voting_model = VotingClassifier(
    estimators=[("lr", lr_pipe), ("rf", rf)],
    voting="soft"
)

models = {
    "Logistic Regression": (lr_pipe, X_train_res, X_test_aligned),
    "Random Forest": (rf, X_train_res, X_test_aligned),
    "XGBoost": (xgb, X_train_res, X_test_aligned),
    "Voting (LR+RF)": (voting_model, X_train_res, X_test_aligned),
}

# ---------- Step 5: 训练与评估 ----------
for name, (model, Xtr, Xte) in models.items():
    model.fit(Xtr, y_train_res)
    preds = model.predict(Xte)
    proba = model.predict_proba(Xte)[:, 1]
    auc = roc_auc_score(y_test, proba)
    print(f"\n{name} AUC: {auc:.4f}")
    print(classification_report(y_test, preds, digits=4))

# ---------- Step 6: SHAP（用已训练好的 XGB） ----------
# 确保用与训练一致的特征空间
xgb.fit(X_train_res, y_train_res)
explainer = shap.TreeExplainer(xgb)
shap_values = explainer.shap_values(X_test_aligned)
shap.summary_plot(shap_values, X_test_aligned, plot_type="bar")

# ---------- Step 7: 蒙特卡洛模拟（场景 A+B+C） ----------
def safe_sample_index(df_bool_mask, frac):
    idx = df_bool_mask[df_bool_mask].index
    if len(idx) == 0:
        return idx  # 空则不抽
    n = max(1, int(len(idx) * frac))
    return np.random.choice(idx, size=n, replace=False)

def monte_carlo_simulation(X_base, trained_model, n_simulations=1000, pA=0.3, pB=0.3, pC=0.2):
    results = []
    cols = set(X_train_res.columns)  # 训练后的列集合

    # 这些列名依赖 get_dummies 后的命名；确保存在才改
    col_contract_one = "Contract_One year"
    col_contract_two = "Contract_Two year"
    col_pay_check = "PaymentMethod_Electronic check"
    col_pay_auto = "PaymentMethod_Credit card (automatic)"
    col_vip = "VIP_Support"
    col_monthly = "MonthlyCharges"

    for _ in range(n_simulations):
        X_sim = X_base.copy()

        # A: 30% 月度合约 -> 一年合约（这里假设原本 "Month-to-month" 列存在）
        if "Contract_Month-to-month" in cols and col_contract_one in cols and col_contract_two in cols:
            month_mask = X_sim["Contract_Month-to-month"] == 1
            month_idx = safe_sample_index(month_mask, pA)
            X_sim.loc[month_idx, "Contract_Month-to-month"] = 0
            X_sim.loc[month_idx, col_contract_one] = 1
            X_sim.loc[month_idx, col_contract_two] = 0

        # B: 30% 电子支票 -> 信用卡自动扣款
        if col_pay_check in cols and col_pay_auto in cols:
            check_mask = X_sim[col_pay_check] == 1
            check_idx = safe_sample_index(check_mask, pB)
            X_sim.loc[check_idx, col_pay_check] = 0
            X_sim.loc[check_idx, col_pay_auto] = 1

        # C: 20% 高账单客户 -> VIP 支持
        if col_monthly in cols and col_vip in cols:
            high_mask = X_sim[col_monthly] > 80
            high_idx = safe_sample_index(high_mask, pC)
            X_sim.loc[high_idx, col_vip] = 1

        # 预测
        proba_sim = trained_model.predict_proba(X_sim)[:, 1]
        churn_rate = np.mean(proba_sim > 0.5)
        results.append(churn_rate)

    return np.mean(results), np.std(results)

mean_churn, std_churn = monte_carlo_simulation(X_test_aligned, voting_model)
print(f"\n蒙特卡洛模拟结果: 平均流失率 {mean_churn:.3f}, 标准差 {std_churn:.3f}")

# ---------- Step 8: ROI 模拟 ----------
def roi_simulation(cost_per_customer, revenue_per_customer, churn_reduction, n_customers=7043):
    saved_customers = n_customers * churn_reduction
    cost = saved_customers * cost_per_customer
    revenue = saved_customers * revenue_per_customer
    roi = (revenue - cost) / max(cost, 1e-9)
    return roi

roi = roi_simulation(cost_per_customer=50, revenue_per_customer=70, churn_reduction=0.05)
print(f"\nROI 模拟结果: ROI={roi*100:.2f}%")

Logistic Regression AUC: 0.8223
              precision    recall  f1-score   support

           0     0.8922    0.7571    0.8191      1552
           1     0.5264    0.7469    0.6175       561

    accuracy                         0.7544      2113
   macro avg     0.7093    0.7520    0.7183      2113
weighted avg     0.7951    0.7544    0.7656      2113


Random Forest AUC: 0.8054
              precision    recall  f1-score   support

           0     0.8836    0.7436    0.8076      1552
           1     0.5068    0.7291    0.5980       561

    accuracy                         0.7397      2113
   macro avg     0.6952    0.7363    0.7028      2113
weighted avg     0.7836    0.7397    0.7519      2113


XGBoost AUC: 0.8291
              precision    recall  f1-score   support

           0     0.8377    0.8711    0.8541      1552
           1     0.5992    0.5330    0.5642       561

    accuracy                         0.7814      2113
   macro avg     0.7184    0.7021    0.7091      2113
weighted avg     0.7744    0.7814    0.7771      2113


Voting (LR+RF) AUC: 0.8184
              precision    recall  f1-score   support

           0     0.8885    0.7597    0.8190      1552
           1     0.5254    0.7362    0.6132       561

    accuracy                         0.7534      2113
   macro avg     0.7070    0.7479    0.7161      2113
weighted avg     0.7921    0.7534    0.7644      2113

蒙特卡洛模拟结果: 平均流失率 0.310, 标准差 0.004

ROI 模拟结果: ROI=40.00%


import pandas as pd
import numpy as np
from statsmodels.stats.proportion import proportions_ztest

# 确保 Churn 是二进制
df["Churn"] = df["Churn"].astype(str).str.strip().replace({"Yes": 1, "No": 0}).astype(int)

# --- Step 1: 直接用原始字段创建标签 ---
df["Contract_label"] = df["Contract"]
df["Payment_label"] = df["PaymentMethod"]

# --- Step 2: 分层抽样随机分组 ---
# 按照 Contract 和 PaymentMethod 分层，保证 Treatment 和 Control 平衡
df["Group"] = "Control"

for _, group in df.groupby(["Contract_label", "Payment_label"]):
    treatment_idx = group.sample(frac=0.5, random_state=42).index
    df.loc[treatment_idx, "Group"] = "Treatment"

print(df["Group"].value_counts())

# --- Step 3: 检查分组分布 ---
print("\n 分组后对比（按合同&支付方式）：")
print(df.groupby(["Contract_label", "Payment_label", "Group"]).size().unstack())

Group
Control      3522
Treatment    3521
Name: count, dtype: int64

 分组后对比（按合同&支付方式）：
Group                                     Control  Treatment
Contract_label Payment_label                                
Month-to-month Bank transfer (automatic)      295        294
               Credit card (automatic)        271        272
               Electronic check               925        925
               Mailed check                   447        446
One year       Bank transfer (automatic)      195        196
               Credit card (automatic)        199        199
               Electronic check               173        174
               Mailed check                   169        168
Two year       Bank transfer (automatic)      282        282
               Credit card (automatic)        291        290
               Electronic check                84         84
               Mailed check                   191        191


# --- Step 1: 导入依赖 ---
import pandas as pd
import numpy as np
from sklearn.model_selection import train_test_split
from sklearn.preprocessing import StandardScaler
from sklearn.linear_model import LogisticRegression
from sklearn.ensemble import RandomForestClassifier, VotingClassifier
from lightgbm import LGBMClassifier
from sklearn.metrics import classification_report, roc_auc_score
from imblearn.over_sampling import SMOTE
import shap
import random

# --- Step 2: 数据准备 ---
if "VIP_Support" not in df.columns:
    df["VIP_Support"] = 0

X = df.drop("Churn", axis=1)
y = df["Churn"].astype(int)

X = X.select_dtypes(include=[np.number])  # 保证数值型
X_train, X_test, y_train, y_test = train_test_split(
    X, y, test_size=0.3, random_state=42, stratify=y
)

scaler = StandardScaler()
X_train_scaled = scaler.fit_transform(X_train)
X_test_scaled = scaler.transform(X_test)

# --- Step 3: Voting 模型 + SMOTE ---
smote = SMOTE(random_state=42)
X_res, y_res = smote.fit_resample(X_train, y_train)

voting_model = VotingClassifier(
    estimators=[
        ("lr", LogisticRegression(max_iter=1000, random_state=42)),
        ("rf", RandomForestClassifier(n_estimators=300, max_depth=15, random_state=42)),
        ("lgbm", LGBMClassifier(n_estimators=300, max_depth=15, random_state=42))
    ],
    voting="soft"
)
voting_model.fit(X_res, y_res)

# --- Step 4: 蒙特卡洛模拟 + uplift 假设 ---
def monte_carlo_with_uplift(X, model, scaler, uplift, threshold=0.5,
                            cost_contract=10, cost_payment=20, cost_vip=15,
                            pct_contract=0.3, pct_payment=0.3, pct_vip=0.2,
                            retention_months=12):
    results = []
    total_customers = len(X)
    avg_monthly_charge = X["MonthlyCharges"].mean()

    for _ in range(500):  # 500次模拟
        X_sim = X.copy()

        # 合约干预
        if "Contract_One year" in X_sim.columns and "Contract_Two year" in X_sim.columns:
            month_idx = X_sim[(X_sim["Contract_One year"] == 0) &
                              (X_sim["Contract_Two year"] == 0)].sample(frac=pct_contract).index
            X_sim.loc[month_idx, "Contract_One year"] = 1

        # 支付干预
        if "PaymentMethod_Electronic check" in X_sim.columns and "PaymentMethod_Credit card (automatic)" in X_sim.columns:
            pay_idx = X_sim[X_sim["PaymentMethod_Electronic check"] == 1].sample(frac=pct_payment).index
            X_sim.loc[pay_idx, "PaymentMethod_Electronic check"] = 0
            X_sim.loc[pay_idx, "PaymentMethod_Credit card (automatic)"] = 1

        # VIP干预
        high_charge_idx = X_sim[X_sim["MonthlyCharges"] > 80].sample(frac=pct_vip).index
        X_sim.loc[high_charge_idx, "VIP_Support"] = 1

        X_sim_scaled = scaler.transform(X_sim)
        proba_sim = model.predict_proba(X_sim_scaled)[:, 1]
        churn_pred = (proba_sim > threshold).astype(int)

        # --- uplift 假设：干预人群中一定比例转化为非流失 ---
        treated_idx = np.random.choice(range(len(X_sim)), size=int(total_customers*(pct_contract+pct_payment+pct_vip)/3), replace=False)
        for idx in treated_idx:
            if churn_pred[idx] == 1 and random.random() < uplift:
                churn_pred[idx] = 0

        results.append(churn_pred.mean())
        
                


    mean_churn = np.mean(results)
    std_churn = np.std(results)

    # ROI 计算
    base_churn = np.mean(model.predict_proba(X_test_scaled)[:, 1] > threshold)
    churn_improvement = (base_churn - mean_churn) * total_customers
    additional_revenue = churn_improvement * avg_monthly_charge * retention_months
    total_cost = (pct_contract * total_customers * cost_contract * retention_months +
                  pct_payment * total_customers * cost_payment +
                  pct_vip * total_customers * cost_vip * retention_months)

    roi = (additional_revenue - total_cost) / total_cost if total_cost > 0 else 0
    return mean_churn, std_churn, roi, additional_revenue, total_cost


# --- Step 5: 敏感性分析（uplift 不同水平） ---
uplift_rates = [0.05, 0.10, 0.20, 0.30]
results_table = []
for uplift in uplift_rates:
    mean_churn, std_churn, roi, rev, cost = monte_carlo_with_uplift(
        X_test, voting_model, scaler, uplift=uplift)
    results_table.append({
        "Uplift": f"{uplift*100:.0f}%",
        "Mean_Churn": mean_churn,
        "Std_Churn": std_churn,
        "ROI": roi,
        "Revenue_Gain": rev,
        "Total_Cost": cost
    })

df_results = pd.DataFrame(results_table)
print("\n 敏感性分析结果：")
print(df_results)

# --- Step 6: SHAP 可解释性 ---
explainer = shap.TreeExplainer(voting_model.named_estimators_["lgbm"])
shap_values = explainer.shap_values(X_test)
shap.summary_plot(shap_values, X_test, max_display=5)

[LightGBM] [Info] Number of positive: 3622, number of negative: 3622
[LightGBM] [Info] Auto-choosing row-wise multi-threading, the overhead of testing was 0.000584 seconds.
You can set `force_row_wise=true` to remove the overhead.
And if memory is not enough, you can set `force_col_wise=true`.
[LightGBM] [Info] Total Bins 330
[LightGBM] [Info] Number of data points in the train set: 7244, number of used features: 3
[LightGBM] [Info] [binary:BoostFromScore]: pavg=0.500000 -> initscore=0.000000

 敏感性分析结果：
  Uplift  Mean_Churn  Std_Churn       ROI  Revenue_Gain  Total_Cost
0     5%    0.373512   0.001575 -0.949561   8312.994486    164814.0
1    10%    0.368591   0.002096 -0.900865  16338.854900    164814.0
2    20%    0.358640   0.002929 -0.802395  32568.104858    164814.0
3    30%    0.348377   0.003581 -0.700834  49306.786234    164814.0


# --- Step 1: 导入依赖 ---
import pandas as pd
import numpy as np
from sklearn.model_selection import train_test_split
from sklearn.preprocessing import StandardScaler
from sklearn.linear_model import LogisticRegression
from sklearn.ensemble import RandomForestClassifier, VotingClassifier
from lightgbm import LGBMClassifier
from sklearn.metrics import classification_report, roc_auc_score
from imblearn.over_sampling import SMOTE
import random

# --- Step 2: 数据准备 ---
# 确保 VIP_Support 存在
if "VIP_Support" not in df.columns:
    df["VIP_Support"] = 0

X = df.drop("Churn", axis=1)
y = df["Churn"].astype(int)

# 确保只保留数值型特征
X = X.select_dtypes(include=[np.number])

X_train, X_test, y_train, y_test = train_test_split(
    X, y, test_size=0.3, random_state=42, stratify=y
)

# 标准化
scaler = StandardScaler()
X_train_scaled = scaler.fit_transform(X_train)
X_test_scaled = scaler.transform(X_test)

# --- Step 3: Voting 模型 + SMOTE ---
smote = SMOTE(random_state=42)
X_res, y_res = smote.fit_resample(X_train, y_train)

voting_model = VotingClassifier(
    estimators=[
        ("lr", LogisticRegression(max_iter=1000, random_state=42)),
        ("rf", RandomForestClassifier(n_estimators=300, max_depth=15, random_state=42)),
        ("lgbm", LGBMClassifier(n_estimators=300, max_depth=15, random_state=42))
    ],
    voting="soft"
)

voting_model.fit(X_res, y_res)

# --- Step 4: 蒙特卡洛模拟与ROI计算 ---
def monte_carlo_simulation_combined(X, model, scaler, n_simulations=1000,
                                   contract_intervention_frac=0.35,
                                   uplift=0.35,
                                   pct_high_risk_intervention=0.35,
                                   threshold=0.5,
                                   avg_monthly_charge=80,
                                   retention_months=8,
                                   pct_intervention=0.35,
                                   cost_per_customer=100):

    results = []
    total_customers = len(X)

    required_contract_cols = {"Contract_One_year", "Contract_Two_year", "Contract_Month_to_month"}
    contract_cols_exist = required_contract_cols.issubset(X.columns)

    for _ in range(n_simulations):
        X_sim = X.copy()

        if contract_cols_exist:
            month_idx = X_sim[(X_sim["Contract_One_year"] == 0) &
                              (X_sim["Contract_Two_year"] == 0) &
                              (X_sim["Contract_Month_to_month"] == 1)].sample(frac=contract_intervention_frac).index
            X_sim.loc[month_idx, "Contract_One_year"] = 1
            X_sim.loc[month_idx, "Contract_Month_to_month"] = 0

        X_sim_scaled = scaler.transform(X_sim)

        proba_sim = model.predict_proba(X_sim_scaled)[:, 1]
        churn_pred = (proba_sim > threshold).astype(int)

        high_risk_count = int(total_customers * pct_high_risk_intervention)
        high_risk_idx = np.argsort(proba_sim)[-high_risk_count:]

        improved_churn = churn_pred.copy()
        for idx in high_risk_idx:
            if churn_pred[idx] == 1:
                if random.random() < uplift:
                    improved_churn[idx] = 0

        churn_rate = improved_churn.mean()
        results.append(churn_rate)

    mean_churn = np.mean(results)
    std_churn = np.std(results)

    base_churn = churn_pred.mean()
    churn_improvement = (base_churn - mean_churn) * total_customers

    additional_revenue = churn_improvement * avg_monthly_charge * retention_months
    total_cost = total_customers * pct_intervention * cost_per_customer
    roi = (additional_revenue - total_cost) / total_cost if total_cost > 0 else 0

    return mean_churn, std_churn, roi, additional_revenue, total_cost

# --- Step 5: 运行模拟 (微调参数进一步降低流失率和ROI) ---
mean_churn, std_churn, roi, revenue_gain, total_cost = monte_carlo_simulation_combined(
    X_test,
    voting_model,
    scaler,
    uplift=0.35,
    pct_intervention=0.35,
    cost_per_customer=100,  # 提高投入成本
    retention_months=8
)

print(f" 蒙特卡洛模拟+uplift结果: 平均流失率 {mean_churn:.3f} (σ={std_churn:.3f})")
print(f" ROI 调整后结果: ROI={roi*100:.2f}%")
print(f"   收益提升=${revenue_gain:,.0f}, 总成本=${total_cost:,.0f}")

[LightGBM] [Info] Number of positive: 3622, number of negative: 3622
[LightGBM] [Info] Auto-choosing col-wise multi-threading, the overhead of testing was 0.000438 seconds.
You can set `force_col_wise=true` to remove the overhead.
[LightGBM] [Info] Total Bins 330
[LightGBM] [Info] Number of data points in the train set: 7244, number of used features: 3
[LightGBM] [Info] [binary:BoostFromScore]: pavg=0.500000 -> initscore=0.000000
 蒙特卡洛模拟+uplift结果: 平均流失率 0.256 (σ=0.006)
 ROI 调整后结果: ROI=123.83%
   收益提升=$165,532, 总成本=$73,955


# --- Step 1: 导入依赖 ---
import pandas as pd
import numpy as np
from sklearn.model_selection import train_test_split
from sklearn.preprocessing import StandardScaler
from sklearn.linear_model import LogisticRegression
from sklearn.ensemble import RandomForestClassifier, VotingClassifier
from lightgbm import LGBMClassifier
from sklearn.metrics import classification_report, roc_auc_score
from imblearn.over_sampling import SMOTE
import random
from scipy.stats import chi2_contingency

# --- Step 2: 数据准备 ---
# 确保 VIP_Support 存在
if "VIP_Support" not in df.columns:
    df["VIP_Support"] = 0

X = df.drop("Churn", axis=1)
y = df["Churn"].astype(int)

# 确保只保留数值型特征
X = X.select_dtypes(include=[np.number])

# 拆分训练和测试集
X_train, X_test, y_train, y_test = train_test_split(
    X, y, test_size=0.3, random_state=42, stratify=y
)

# 标准化
scaler = StandardScaler()
X_train_scaled = scaler.fit_transform(X_train)
X_test_scaled = scaler.transform(X_test)

# --- Step 3: Voting 模型 + SMOTE ---
smote = SMOTE(random_state=42)
X_res, y_res = smote.fit_resample(X_train, y_train)

voting_model = VotingClassifier(
    estimators=[
        ("lr", LogisticRegression(max_iter=1000, random_state=42)),
        ("rf", RandomForestClassifier(n_estimators=300, max_depth=15, random_state=42)),
        ("lgbm", LGBMClassifier(n_estimators=300, max_depth=15, random_state=42))
    ],
    voting="soft"
)

voting_model.fit(X_res, y_res)

# --- Step 4: 蒙特卡洛模拟与ROI计算 ---
def monte_carlo_simulation_combined(X, model, scaler, n_simulations=1000,
                                   contract_intervention_frac=0.35,
                                   uplift=0.35,
                                   pct_high_risk_intervention=0.35,
                                   threshold=0.5,
                                   avg_monthly_charge=80,
                                   retention_months=8,
                                   pct_intervention=0.35,
                                   cost_per_customer=100):

    results = []
    total_customers = len(X)

    for _ in range(n_simulations):
        X_sim = X.copy()

        # 标准化
        X_sim_scaled = scaler.transform(X_sim)

        # 模型预测
        proba_sim = model.predict_proba(X_sim_scaled)[:, 1]
        churn_pred = (proba_sim > threshold).astype(int)

        # 找到高风险客户
        high_risk_count = int(total_customers * pct_high_risk_intervention)
        high_risk_idx = np.argsort(proba_sim)[-high_risk_count:]

        # 应用uplift假设
        improved_churn = churn_pred.copy()
        for idx in high_risk_idx:
            if churn_pred[idx] == 1:
                if random.random() < uplift:
                    improved_churn[idx] = 0

        churn_rate = improved_churn.mean()
        results.append(churn_rate)

    mean_churn = np.mean(results)
    std_churn = np.std(results)

    base_churn = churn_pred.mean()
    churn_improvement = (base_churn - mean_churn) * total_customers

    additional_revenue = churn_improvement * avg_monthly_charge * retention_months
    total_cost = total_customers * pct_intervention * cost_per_customer
    roi = (additional_revenue - total_cost) / total_cost if total_cost > 0 else 0

    return mean_churn, std_churn, roi, additional_revenue, total_cost

# --- Step 5: 运行模拟 ---
mean_churn, std_churn, roi, revenue_gain, total_cost = monte_carlo_simulation_combined(
    X_test,
    voting_model,
    scaler,
    uplift=0.35,
    pct_intervention=0.35,
    cost_per_customer=100,  # 每个干预客户成本
    retention_months=8
)

print(f" 蒙特卡洛模拟+uplift结果: 平均流失率 {mean_churn:.3f} (σ={std_churn:.3f})")
print(f" ROI 调整后结果: ROI={roi*100:.2f}%")
print(f"   收益提升=${revenue_gain:,.0f}, 总成本=${total_cost:,.0f}")

# --- Step 6: AB测试模块 ---
def ab_test_simulation(X, model, scaler, uplift=0.35, threshold=0.5, test_frac=0.5):
    X_scaled = scaler.transform(X)
    proba = model.predict_proba(X_scaled)[:, 1]
    churn_pred = (proba > threshold).astype(int)

    # 随机分组
    np.random.seed(42)
    treatment_idx = np.random.choice(X.index, size=int(len(X)*test_frac), replace=False)
    control_idx = X.index.difference(treatment_idx)

    df_ab = pd.DataFrame({
        "Group": ["Treatment" if i in treatment_idx else "Control" for i in X.index],
        "Churn_pred": churn_pred
    })

    # Treatment 组应用 uplift 假设
    improved_churn = df_ab.loc[df_ab["Group"]=="Treatment", "Churn_pred"].copy()
    for i in improved_churn.index:
        if improved_churn[i] == 1 and random.random() < uplift:
            improved_churn[i] = 0
    df_ab.loc[df_ab["Group"]=="Treatment", "Churn_pred"] = improved_churn

    # 流失率统计
    churn_rates = df_ab.groupby("Group")["Churn_pred"].mean()

    # 卡方检验
    contingency_table = pd.crosstab(df_ab["Group"], df_ab["Churn_pred"])
    chi2, p_value, _, _ = chi2_contingency(contingency_table)

    return churn_rates, chi2, p_value

# --- Step 7: 执行 AB 测试 ---
churn_rates, chi2, p_value = ab_test_simulation(X_test, voting_model, scaler, uplift=0.35)

print("\n AB测试结果")
print(churn_rates)
print(f" 卡方检验: Chi2={chi2:.3f}, p-value={p_value:.4f}")
if p_value < 0.05:
    print(" 差异显著，干预有效")
else:
    print(" 差异不显著，需要更多样本或优化策略")

[LightGBM] [Info] Number of positive: 3622, number of negative: 3622
[LightGBM] [Info] Auto-choosing col-wise multi-threading, the overhead of testing was 0.000414 seconds.
You can set `force_col_wise=true` to remove the overhead.
[LightGBM] [Info] Total Bins 330
[LightGBM] [Info] Number of data points in the train set: 7244, number of used features: 3
[LightGBM] [Info] [binary:BoostFromScore]: pavg=0.500000 -> initscore=0.000000
 蒙特卡洛模拟+uplift结果: 平均流失率 0.256 (σ=0.006)
 ROI 调整后结果: ROI=123.72%
   收益提升=$165,453, 总成本=$73,955

 AB测试结果
Group
Control      0.378430
Treatment    0.245265
Name: Churn_pred, dtype: float64
 卡方检验: Chi2=43.030, p-value=0.0000
 差异显著，干预有效


# --- Step 8: Precision-Recall 曲线可视化 ---
from sklearn.metrics import precision_recall_curve, auc
import matplotlib.pyplot as plt

def plot_pr_curve(X, y, model, scaler):
    X_scaled = scaler.transform(X)
    y_proba = model.predict_proba(X_scaled)[:, 1]
    precision, recall, thresholds = precision_recall_curve(y, y_proba)
    pr_auc = auc(recall, precision)

    plt.figure(figsize=(7,5))
    plt.plot(recall, precision, linewidth=2)
    plt.xlabel("Recall")
    plt.ylabel("Precision")
    plt.title(f"Precision-Recall Curve (PR-AUC = {pr_auc:.3f})")
    plt.grid(True)
    plt.show()

# 调用
plot_pr_curve(X_test, y_test, voting_model, scaler)

# --- Step 9: ROI 成本拆分表 ---
def roi_cost_breakdown(total_customers, churn_rate, reduced_churn_rate, 
                       avg_revenue_per_customer=1000, intervention_cost_per_customer=100, 
                       high_risk_ratio=0.35, retention_months=8):
    baseline_churn_customers = total_customers * churn_rate
    baseline_loss = baseline_churn_customers * avg_revenue_per_customer

    churn_customers_after = total_customers * reduced_churn_rate
    loss_after = churn_customers_after * avg_revenue_per_customer

    target_customers = int(total_customers * high_risk_ratio)
    intervention_cost = target_customers * intervention_cost_per_customer

    revenue_gain = baseline_loss - loss_after
    roi = (revenue_gain - intervention_cost) / intervention_cost * 100 if intervention_cost > 0 else 0

    roi_df = pd.DataFrame({
        "Metric": ["Total Customers", "Baseline Churn Customers", "Churn After Intervention",
                   "Revenue Loss (Baseline)", "Revenue Loss (After)", "Revenue Gain", 
                   "Intervention Cost", "ROI (%)"],
        "Value": [total_customers, baseline_churn_customers, churn_customers_after,
                  baseline_loss, loss_after, revenue_gain, intervention_cost, roi]
    })
    return roi_df

# 调用（用Monte Carlo结果替换流失率）
roi_table = roi_cost_breakdown(
    total_customers=len(X_test),
    churn_rate=0.228,
    reduced_churn_rate=mean_churn,
    avg_revenue_per_customer=100*8,  # 月费 * 预计留存月数
    intervention_cost_per_customer=100,
    high_risk_ratio=0.35,
    retention_months=8
)
print("\n ROI 成本拆分表")
print(roi_table)

# --- Step 10: AB测试样本量计算 ---
from statsmodels.stats.power import NormalIndPower
import math

def ab_test_sample_size(alpha=0.05, power=0.8, baseline_rate=0.228, expected_rate=0.169):
    p1, p2 = baseline_rate, expected_rate
    cohen_h = 2 * math.asin(math.sqrt(p1)) - 2 * math.asin(math.sqrt(p2))
    analysis = NormalIndPower()
    sample_size_per_group = analysis.solve_power(effect_size=cohen_h, alpha=alpha, power=power, ratio=1, alternative='two-sided')
    total_sample_size = math.ceil(sample_size_per_group * 2)

    df_sample = pd.DataFrame({
        "Parameter": ["Alpha", "Power", "Baseline Rate", "Expected Rate", "Effect Size (Cohen's h)", 
                      "Sample Size per Group", "Total Sample Size"],
        "Value": [alpha, power, baseline_rate, expected_rate, cohen_h, sample_size_per_group, total_sample_size]
    })
    return df_sample

# 调用
sample_size_table = ab_test_sample_size()
print("\n AB测试样本量计算表")
print(sample_size_table)

 ROI 成本拆分表
                     Metric          Value
0           Total Customers    2113.000000
1  Baseline Churn Customers     481.764000
2  Churn After Intervention     541.479000
3   Revenue Loss (Baseline)  385411.200000
4      Revenue Loss (After)  433183.200000
5              Revenue Gain  -47772.000000
6         Intervention Cost   73900.000000
7                   ROI (%)    -164.644114

 AB测试样本量计算表
                 Parameter        Value
0                    Alpha     0.050000
1                    Power     0.800000
2            Baseline Rate     0.228000
3            Expected Rate     0.169000
4  Effect Size (Cohen's h)     0.148287
5    Sample Size per Group   713.886878
6        Total Sample Size  1428.000000


from sklearn.model_selection import cross_val_score
scores = cross_val_score(voting_model, X_res, y_res, cv=5, scoring="average_precision")
print(f"5折交叉验证的平均PR-AUC: {scores.mean():.3f} ± {scores.std():.3f}")

[LightGBM] [Warning] Found whitespace in feature_names, replace with underlines
[LightGBM] [Info] Number of positive: 2898, number of negative: 2897
[LightGBM] [Info] Auto-choosing row-wise multi-threading, the overhead of testing was 0.001304 seconds.
You can set `force_row_wise=true` to remove the overhead.
And if memory is not enough, you can set `force_col_wise=true`.
[LightGBM] [Info] Total Bins 625
[LightGBM] [Info] Number of data points in the train set: 5795, number of used features: 24
[LightGBM] [Info] [binary:BoostFromScore]: pavg=0.500086 -> initscore=0.000345
[LightGBM] [Info] Start training from score 0.000345
[LightGBM] [Warning] Found whitespace in feature_names, replace with underlines
[LightGBM] [Info] Number of positive: 2898, number of negative: 2897
[LightGBM] [Info] Auto-choosing row-wise multi-threading, the overhead of testing was 0.000920 seconds.
You can set `force_row_wise=true` to remove the overhead.
And if memory is not enough, you can set `force_col_wise=true`.
[LightGBM] [Info] Total Bins 625
[LightGBM] [Info] Number of data points in the train set: 5795, number of used features: 24
[LightGBM] [Info] [binary:BoostFromScore]: pavg=0.500086 -> initscore=0.000345
[LightGBM] [Info] Start training from score 0.000345
[LightGBM] [Warning] Found whitespace in feature_names, replace with underlines
[LightGBM] [Info] Number of positive: 2897, number of negative: 2898
[LightGBM] [Info] Auto-choosing row-wise multi-threading, the overhead of testing was 0.004234 seconds.
You can set `force_row_wise=true` to remove the overhead.
And if memory is not enough, you can set `force_col_wise=true`.
[LightGBM] [Info] Total Bins 625
[LightGBM] [Info] Number of data points in the train set: 5795, number of used features: 24
[LightGBM] [Info] [binary:BoostFromScore]: pavg=0.499914 -> initscore=-0.000345
[LightGBM] [Info] Start training from score -0.000345
[LightGBM] [Warning] Found whitespace in feature_names, replace with underlines
[LightGBM] [Info] Number of positive: 2897, number of negative: 2898
[LightGBM] [Info] Auto-choosing col-wise multi-threading, the overhead of testing was 0.001052 seconds.
You can set `force_col_wise=true` to remove the overhead.
[LightGBM] [Info] Total Bins 625
[LightGBM] [Info] Number of data points in the train set: 5795, number of used features: 24
[LightGBM] [Info] [binary:BoostFromScore]: pavg=0.499914 -> initscore=-0.000345
[LightGBM] [Info] Start training from score -0.000345
[LightGBM] [Warning] Found whitespace in feature_names, replace with underlines
[LightGBM] [Info] Number of positive: 2898, number of negative: 2898
[LightGBM] [Info] Auto-choosing row-wise multi-threading, the overhead of testing was 0.000872 seconds.
You can set `force_row_wise=true` to remove the overhead.
And if memory is not enough, you can set `force_col_wise=true`.
[LightGBM] [Info] Total Bins 625
[LightGBM] [Info] Number of data points in the train set: 5796, number of used features: 24
[LightGBM] [Info] [binary:BoostFromScore]: pavg=0.500000 -> initscore=0.000000
5折交叉验证的平均PR-AUC: 0.929 ± 0.062


# --- Step 8: Precision-Recall 曲线 + 5折交叉验证 ---
from sklearn.metrics import precision_recall_curve, auc
from sklearn.model_selection import cross_val_score
import matplotlib.pyplot as plt

def plot_pr_curve_and_cv(X_train, y_train, X_test, y_test, model, scaler):
    # 交叉验证 PR-AUC
    scores = cross_val_score(model, X_train, y_train, cv=5, scoring="average_precision")
    print(f"5折交叉验证的平均PR-AUC: {scores.mean():.3f} ± {scores.std():.3f}")

    # 绘制PR曲线
    X_scaled = scaler.transform(X_test)
    y_proba = model.predict_proba(X_scaled)[:, 1]
    precision, recall, thresholds = precision_recall_curve(y_test, y_proba)
    pr_auc = auc(recall, precision)

    plt.figure(figsize=(7,5))
    plt.plot(recall, precision, linewidth=2)
    plt.xlabel("Recall")
    plt.ylabel("Precision")
    plt.title(f"Precision-Recall Curve (PR-AUC = {pr_auc:.3f})")
    plt.grid(True)
    plt.show()

# 调用
plot_pr_curve_and_cv(X_train, y_train, X_test, y_test, voting_model, scaler)


## --- 改进版 ROI 成本拆分（每个Metric都有解释） ---
def roi_cost_breakdown_aligned(total_customers, base_churn_rate, mean_churn_rate,
                               avg_monthly_charge=80, retention_months=8,
                               intervention_cost_per_customer=100,
                               high_risk_ratio=0.35):

    baseline_churn_customers = total_customers * base_churn_rate
    churn_customers_after = total_customers * mean_churn_rate

    avg_revenue_per_customer = avg_monthly_charge * retention_months
    baseline_loss = baseline_churn_customers * avg_revenue_per_customer
    loss_after = churn_customers_after * avg_revenue_per_customer

    target_customers = int(total_customers * high_risk_ratio)
    intervention_cost = target_customers * intervention_cost_per_customer

    revenue_gain = baseline_loss - loss_after
    roi = (revenue_gain - intervention_cost) / intervention_cost * 100 if intervention_cost > 0 else 0

    roi_df = pd.DataFrame({
        "Metric": [
            "Total Customers",
            "Baseline Churn Customers",
            "Churn After Intervention",
            "Revenue Loss (Baseline)",
            "Revenue Loss (After)",
            "Revenue Gain",
            "Intervention Cost",
            "ROI (%)"
        ],
        "Value": [
            total_customers,
            baseline_churn_customers,
            churn_customers_after,
            baseline_loss,
            loss_after,
            revenue_gain,
            intervention_cost,
            roi
        ],
        "Cost Breakdown": [
            "分析样本中的总客户数",
            "按基线流失率估算的流失客户数",
            "干预后预测的流失客户数",
            "流失客户导致的收入损失（基线）",
            "干预后流失客户导致的收入损失",
            "干预带来的收入损失减少（即收益提升）",
            "干预投入成本（短信、优惠、人工等）",
            "收益相对干预成本的比例"
        ]
    })
    return roi_df

# --- 调用 ---
roi_table_aligned = roi_cost_breakdown_aligned(
    total_customers=len(X_test),
    base_churn_rate=0.228,
    mean_churn_rate=mean_churn,
    avg_monthly_charge=80,
    retention_months=8,
    intervention_cost_per_customer=100,
    high_risk_ratio=0.35
)

print("\n 改进版 ROI 成本拆分表")
print(roi_table_aligned)

# --- Step 10: AB测试样本量计算 ---
from statsmodels.stats.power import NormalIndPower
import math

def ab_test_sample_size(alpha=0.05, power=0.8, baseline_rate=0.228, expected_rate=0.169):
    p1, p2 = baseline_rate, expected_rate
    cohen_h = 2 * math.asin(math.sqrt(p1)) - 2 * math.asin(math.sqrt(p2))
    analysis = NormalIndPower()
    sample_size_per_group = analysis.solve_power(effect_size=cohen_h, alpha=alpha, power=power, ratio=1, alternative='two-sided')
    total_sample_size = math.ceil(sample_size_per_group * 2)

    df_sample = pd.DataFrame({
        "Parameter": ["Alpha", "Power", "Baseline Rate", "Expected Rate", "Effect Size (Cohen's h)", 
                      "Sample Size per Group", "Total Sample Size"],
        "Value": [alpha, power, baseline_rate, expected_rate, cohen_h, sample_size_per_group, total_sample_size]
    })
    return df_sample

# 调用
sample_size_table = ab_test_sample_size()
print("\n AB测试样本量计算表")
print(sample_size_table)

[LightGBM] [Info] Number of positive: 1047, number of negative: 2897
[LightGBM] [Info] Auto-choosing row-wise multi-threading, the overhead of testing was 0.000121 seconds.
You can set `force_row_wise=true` to remove the overhead.
And if memory is not enough, you can set `force_col_wise=true`.
[LightGBM] [Info] Total Bins 330
[LightGBM] [Info] Number of data points in the train set: 3944, number of used features: 3
[LightGBM] [Info] [binary:BoostFromScore]: pavg=0.265467 -> initscore=-1.017747
[LightGBM] [Info] Start training from score -1.017747
[LightGBM] [Warning] No further splits with positive gain, best gain: -inf
[LightGBM] [Info] Number of positive: 1047, number of negative: 2897
[LightGBM] [Info] Auto-choosing col-wise multi-threading, the overhead of testing was 0.000312 seconds.
You can set `force_col_wise=true` to remove the overhead.
[LightGBM] [Info] Total Bins 330
[LightGBM] [Info] Number of data points in the train set: 3944, number of used features: 3
[LightGBM] [Info] [binary:BoostFromScore]: pavg=0.265467 -> initscore=-1.017747
[LightGBM] [Info] Start training from score -1.017747
[LightGBM] [Warning] No further splits with positive gain, best gain: -inf
[LightGBM] [Warning] No further splits with positive gain, best gain: -inf
[LightGBM] [Warning] No further splits with positive gain, best gain: -inf
[LightGBM] [Warning] No further splits with positive gain, best gain: -inf
[LightGBM] [Warning] No further splits with positive gain, best gain: -inf
[LightGBM] [Info] Number of positive: 1046, number of negative: 2898
[LightGBM] [Info] Auto-choosing row-wise multi-threading, the overhead of testing was 0.000078 seconds.
You can set `force_row_wise=true` to remove the overhead.
And if memory is not enough, you can set `force_col_wise=true`.
[LightGBM] [Info] Total Bins 330
[LightGBM] [Info] Number of data points in the train set: 3944, number of used features: 3
[LightGBM] [Info] [binary:BoostFromScore]: pavg=0.265213 -> initscore=-1.019047
[LightGBM] [Info] Start training from score -1.019047
[LightGBM] [Warning] No further splits with positive gain, best gain: -inf
[LightGBM] [Warning] No further splits with positive gain, best gain: -inf
[LightGBM] [Warning] No further splits with positive gain, best gain: -inf
[LightGBM] [Warning] No further splits with positive gain, best gain: -inf
[LightGBM] [Warning] No further splits with positive gain, best gain: -inf
[LightGBM] [Warning] No further splits with positive gain, best gain: -inf
[LightGBM] [Info] Number of positive: 1046, number of negative: 2898
[LightGBM] [Info] Auto-choosing col-wise multi-threading, the overhead of testing was 0.000331 seconds.
You can set `force_col_wise=true` to remove the overhead.
[LightGBM] [Info] Total Bins 330
[LightGBM] [Info] Number of data points in the train set: 3944, number of used features: 3
[LightGBM] [Info] [binary:BoostFromScore]: pavg=0.265213 -> initscore=-1.019047
[LightGBM] [Info] Start training from score -1.019047
[LightGBM] [Warning] No further splits with positive gain, best gain: -inf
[LightGBM] [Warning] No further splits with positive gain, best gain: -inf
[LightGBM] [Info] Number of positive: 1046, number of negative: 2898
[LightGBM] [Info] Auto-choosing row-wise multi-threading, the overhead of testing was 0.000117 seconds.
You can set `force_row_wise=true` to remove the overhead.
And if memory is not enough, you can set `force_col_wise=true`.
[LightGBM] [Info] Total Bins 330
[LightGBM] [Info] Number of data points in the train set: 3944, number of used features: 3
[LightGBM] [Info] [binary:BoostFromScore]: pavg=0.265213 -> initscore=-1.019047
[LightGBM] [Info] Start training from score -1.019047
[LightGBM] [Warning] No further splits with positive gain, best gain: -inf
[LightGBM] [Warning] No further splits with positive gain, best gain: -inf
5折交叉验证的平均PR-AUC: 0.610 ± 0.030

 改进版 ROI 成本拆分表
                     Metric          Value      Cost Breakdown
0           Total Customers    2113.000000          分析样本中的总客户数
1  Baseline Churn Customers     481.764000      按基线流失率估算的流失客户数
2  Churn After Intervention     541.479000         干预后预测的流失客户数
3   Revenue Loss (Baseline)  308328.960000     流失客户导致的收入损失（基线）
4      Revenue Loss (After)  346546.560000      干预后流失客户导致的收入损失
5              Revenue Gain  -38217.600000  干预带来的收入损失减少（即收益提升）
6         Intervention Cost   73900.000000   干预投入成本（短信、优惠、人工等）
7                   ROI (%)    -151.715291         收益相对干预成本的比例

 AB测试样本量计算表
                 Parameter        Value
0                    Alpha     0.050000
1                    Power     0.800000
2            Baseline Rate     0.228000
3            Expected Rate     0.169000
4  Effect Size (Cohen's h)     0.148287
5    Sample Size per Group   713.886878
6        Total Sample Size  1428.000000


# --- Step 11: 业务迁移模块 ---
def business_case_migration(X, y, model, scaler, scenario="banking",
                            uplift=0.3, cost_per_customer=120,
                            avg_monthly_charge=100, retention_months=12,
                            high_risk_ratio=0.3, test_frac=0.5):

    """
    功能: 将现有 churn 框架迁移到不同行业 (银行/订阅/SaaS)
    
    参数:
    - scenario: "banking", "subscription", "saas"
    - uplift: 干预成功率 (默认 30%)
    - cost_per_customer: 干预成本
    - avg_monthly_charge: 每客户平均月收入 (ARPU)
    - retention_months: 平均留存期
    - high_risk_ratio: 干预客户比例
    - test_frac: AB 测试对照/实验比例
    """

    # 行业参数设置
    if scenario == "banking":
        uplift = 0.4
        avg_monthly_charge = 150  # 银行信用卡 ARPU 较高
        retention_months = 18
        cost_per_customer = 200   # 干预成本更高（专员电话、积分）
    elif scenario == "subscription":
        uplift = 0.25
        avg_monthly_charge = 15   # 流媒体订阅月费
        retention_months = 6
        cost_per_customer = 30    # 低成本（优惠券/推送）
    elif scenario == "saas":
        uplift = 0.35
        avg_monthly_charge = 60   # SaaS 软件订阅
        retention_months = 12
        cost_per_customer = 120   # 客服干预成本
    
    # 标准化
    X_scaled = scaler.transform(X)
    proba = model.predict_proba(X_scaled)[:, 1]
    churn_pred = (proba > 0.5).astype(int)
    total_customers = len(X)
    baseline_rate = churn_pred.mean()

    # 高风险干预
    high_risk_count = int(total_customers * high_risk_ratio)
    high_risk_idx = np.argsort(proba)[-high_risk_count:]
    improved_churn = churn_pred.copy()
    for idx in high_risk_idx:
        if churn_pred[idx] == 1 and random.random() < uplift:
            improved_churn[idx] = 0
    
    # 新流失率
    mean_churn_rate = improved_churn.mean()

    # ROI 成本拆分
    roi_table = roi_cost_breakdown_aligned(
        total_customers=total_customers,
        base_churn_rate=baseline_rate,
        mean_churn_rate=mean_churn_rate,
        avg_monthly_charge=avg_monthly_charge,
        retention_months=retention_months,
        intervention_cost_per_customer=cost_per_customer,
        high_risk_ratio=high_risk_ratio
    )

    print(f"\n {scenario.upper()} 迁移结果")
    print(roi_table)

    return roi_table


# --- 调用示例 ---
# 银行业务
banking_results = business_case_migration(X_test, y_test, voting_model, scaler, scenario="banking")

# 订阅服务
subscription_results = business_case_migration(X_test, y_test, voting_model, scaler, scenario="subscription")

# SaaS 产品
saas_results = business_case_migration(X_test, y_test, voting_model, scaler, scenario="saas")

 BANKING 迁移结果
                     Metric         Value      Cost Breakdown
0           Total Customers  2.113000e+03          分析样本中的总客户数
1  Baseline Churn Customers  8.000000e+02      按基线流失率估算的流失客户数
2  Churn After Intervention  5.560000e+02         干预后预测的流失客户数
3   Revenue Loss (Baseline)  2.160000e+06     流失客户导致的收入损失（基线）
4      Revenue Loss (After)  1.501200e+06      干预后流失客户导致的收入损失
5              Revenue Gain  6.588000e+05  干预带来的收入损失减少（即收益提升）
6         Intervention Cost  1.266000e+05   干预投入成本（短信、优惠、人工等）
7                   ROI (%)  4.203791e+02         收益相对干预成本的比例

 SUBSCRIPTION 迁移结果
                     Metric         Value      Cost Breakdown
0           Total Customers   2113.000000          分析样本中的总客户数
1  Baseline Churn Customers    800.000000      按基线流失率估算的流失客户数
2  Churn After Intervention    645.000000         干预后预测的流失客户数
3   Revenue Loss (Baseline)  72000.000000     流失客户导致的收入损失（基线）
4      Revenue Loss (After)  58050.000000      干预后流失客户导致的收入损失
5              Revenue Gain  13950.000000  干预带来的收入损失减少（即收益提升）
6         Intervention Cost  18990.000000   干预投入成本（短信、优惠、人工等）
7                   ROI (%)    -26.540284         收益相对干预成本的比例

 SAAS 迁移结果
                     Metric          Value      Cost Breakdown
0           Total Customers    2113.000000          分析样本中的总客户数
1  Baseline Churn Customers     800.000000      按基线流失率估算的流失客户数
2  Churn After Intervention     586.000000         干预后预测的流失客户数
3   Revenue Loss (Baseline)  576000.000000     流失客户导致的收入损失（基线）
4      Revenue Loss (After)  421920.000000      干预后流失客户导致的收入损失
5              Revenue Gain  154080.000000  干预带来的收入损失减少（即收益提升）
6         Intervention Cost   75960.000000   干预投入成本（短信、优惠、人工等）
7                   ROI (%)     102.843602         收益相对干预成本的比例

Number of variables	31
Number of observations	7043
Missing cells	0
Missing cells (%)	0.0%
Duplicate rows	20
Duplicate rows (%)	0.3%
Total size in memory	694.8 KiB
Average record size in memory	101.0 B

Categorical	7
Numeric	3
Boolean	21

Dataset has 20 (0.3%) duplicate rows	Duplicates
`Contract_Two year` is highly overall correlated with `tenure`	High correlation
`DeviceProtection_No internet service` is highly overall correlated with `InternetService_No` and 6 other fields	High correlation
`DeviceProtection_Yes` is highly overall correlated with `MonthlyCharges` and 1 other fields	High correlation
`InternetService_Fiber optic` is highly overall correlated with `MonthlyCharges`	High correlation
`InternetService_No` is highly overall correlated with `DeviceProtection_No internet service` and 6 other fields	High correlation
`MonthlyCharges` is highly overall correlated with `DeviceProtection_No internet service` and 14 other fields	High correlation
`MultipleLines_No phone service` is highly overall correlated with `MonthlyCharges` and 1 other fields	High correlation
`MultipleLines_Yes` is highly overall correlated with `MonthlyCharges`	High correlation
`OnlineBackup_No internet service` is highly overall correlated with `DeviceProtection_No internet service` and 6 other fields	High correlation
`OnlineBackup_Yes` is highly overall correlated with `TotalCharges`	High correlation
`OnlineSecurity_No internet service` is highly overall correlated with `DeviceProtection_No internet service` and 6 other fields	High correlation
`PhoneService` is highly overall correlated with `MonthlyCharges` and 1 other fields	High correlation
`StreamingMovies_No internet service` is highly overall correlated with `DeviceProtection_No internet service` and 6 other fields	High correlation
`StreamingMovies_Yes` is highly overall correlated with `MonthlyCharges` and 2 other fields	High correlation
`StreamingTV_No internet service` is highly overall correlated with `DeviceProtection_No internet service` and 6 other fields	High correlation
`StreamingTV_Yes` is highly overall correlated with `MonthlyCharges` and 2 other fields	High correlation
`TechSupport_No internet service` is highly overall correlated with `DeviceProtection_No internet service` and 6 other fields	High correlation
`TotalCharges` is highly overall correlated with `DeviceProtection_Yes` and 5 other fields	High correlation
`tenure` is highly overall correlated with `Contract_Two year` and 1 other fields	High correlation
`PhoneService` is highly imbalanced (54.1%)	Imbalance
`MultipleLines_No phone service` is highly imbalanced (54.1%)	Imbalance

Analysis started	2025-08-09 10:12:53.387535
Analysis finished	2025-08-09 10:12:56.064012
Duration	2.68 seconds
Software version	ydata-profiling vv4.16.1
Download configuration	config.json

Distinct	2
Distinct (%)	< 0.1%
Missing	0
Missing (%)	0.0%
Memory size	399.0 KiB

Total characters	7043
Distinct characters	2
Distinct categories	1 ?
Distinct scripts	1 ?
Distinct blocks	1 ?

Distinct	73
Distinct (%)	1.0%
Missing	0
Missing (%)	0.0%
Infinite	0
Infinite (%)	0.0%
Mean	32.371149

Minimum	0
Maximum	72
Zeros	11
Zeros (%)	0.2%
Negative	0
Negative (%)	0.0%
Memory size	55.1 KiB

Minimum	0
5-th percentile	1
Q1	9
median	29
Q3	55
95-th percentile	72
Maximum	72
Range	72
Interquartile range (IQR)	46

Standard deviation	24.559481
Coefficient of variation (CV)	0.75868426
Kurtosis	-1.3873716
Mean	32.371149
Median Absolute Deviation (MAD)	22
Skewness	0.23953975
Sum	227990
Variance	603.16811
Monotonicity	Not monotonic

Value	Count	Frequency (%)
1	613	8.7%
72	362	5.1%
2	238	3.4%
3	200	2.8%
4	176	2.5%
71	170	2.4%
5	133	1.9%
7	131	1.9%
8	123	1.7%
70	119	1.7%
Other values (63)	4778	67.8%

Distinct	1585
Distinct (%)	22.5%
Missing	0
Missing (%)	0.0%
Infinite	0
Infinite (%)	0.0%
Mean	64.761692

Minimum	18.25
Maximum	118.75
Zeros	0
Zeros (%)	0.0%
Negative	0
Negative (%)	0.0%
Memory size	55.1 KiB

Standard deviation	30.090047
Coefficient of variation (CV)	0.46462725
Kurtosis	-1.2572597
Mean	64.761692
Median Absolute Deviation (MAD)	24.05
Skewness	-0.22052443
Sum	456116.6
Variance	905.41093
Monotonicity	Not monotonic

Distinct	6531
Distinct (%)	92.7%
Missing	0
Missing (%)	0.0%
Infinite	0
Infinite (%)	0.0%
Mean	2279.7343

1	3555
0	3488

Max length	1
Median length	1
Mean length	1
Min length	1

1st row	0
2nd row	1
3rd row	1
4th row	1
5th row	0

Value	Count	Frequency (%)
1	3555	50.5%
0	3488	49.5%

Minimum	0
Maximum	8684.8
Zeros	11
Zeros (%)	0.2%
Negative	0
Negative (%)	0.0%
Memory size	55.1 KiB

Minimum	0
5-th percentile	48.6
Q1	398.55
median	1394.55
Q3	3786.6
95-th percentile	6921.025
Maximum	8684.8
Range	8684.8
Interquartile range (IQR)	3388.05

Standard deviation	2266.7945
Coefficient of variation (CV)	0.99432397
Kurtosis	-0.22857981
Mean	2279.7343
Median Absolute Deviation (MAD)	1223.1
Skewness	0.96323465
Sum	16056169
Variance	5138357.2
Monotonicity	Not monotonic

	Churn	Contract_One year	Contract_Two year	Dependents	DeviceProtection_No internet service	DeviceProtection_Yes	InternetService_Fiber optic	InternetService_No	MonthlyCharges	MultipleLines_No phone service	MultipleLines_Yes	OnlineBackup_No internet service	OnlineBackup_Yes	OnlineSecurity_No internet service	OnlineSecurity_Yes	PaperlessBilling	Partner	PaymentMethod_Credit card (automatic)	PaymentMethod_Electronic check	PaymentMethod_Mailed check	PhoneService	SeniorCitizen	StreamingMovies_No internet service	StreamingMovies_Yes	StreamingTV_No internet service	StreamingTV_Yes	TechSupport_No internet service	TechSupport_Yes	TotalCharges	gender	tenure
Churn	1.000	0.177	0.302	0.163	0.227	0.065	0.307	0.227	0.276	0.000	0.038	0.227	0.081	0.227	0.170	0.191	0.150	0.133	0.301	0.091	0.000	0.150	0.227	0.060	0.227	0.062	0.227	0.164	0.213	0.000	0.364
Contract_One year	0.177	1.000	0.289	0.067	0.036	0.101	0.075	0.036	0.128	0.000	0.000	0.036	0.083	0.036	0.099	0.050	0.082	0.066	0.108	0.000	0.000	0.044	0.036	0.063	0.036	0.060	0.036	0.095	0.208	0.000	0.307
Contract_Two year	0.302	0.289	1.000	0.204	0.218	0.164	0.211	0.218	0.303	0.000	0.105	0.218	0.110	0.218	0.191	0.147	0.247	0.172	0.282	0.000	0.000	0.116	0.218	0.073	0.218	0.071	0.218	0.240	0.396	0.000	0.596
Dependents	0.163	0.067	0.204	1.000	0.139	0.007	0.165	0.139	0.142	0.000	0.021	0.139	0.020	0.139	0.080	0.110	0.452	0.059	0.150	0.057	0.000	0.210	0.139	0.038	0.139	0.011	0.139	0.062	0.091	0.000	0.163
DeviceProtection_No internet service	0.227	0.036	0.218	0.139	1.000	0.380	0.465	1.000	0.967	0.171	0.210	1.000	0.381	1.000	0.333	0.320	0.000	0.000	0.284	0.321	0.171	0.182	1.000	0.418	1.000	0.415	1.000	0.336	0.429	0.000	0.023
DeviceProtection_Yes	0.065	0.101	0.164	0.007	0.380	1.000	0.175	0.380	0.519	0.070	0.200	0.380	0.303	0.380	0.275	0.103	0.153	0.111	0.000	0.187	0.070	0.058	0.380	0.402	0.380	0.390	0.380	0.333	0.522	0.000	0.358
InternetService_Fiber optic	0.307	0.075	0.211	0.165	0.465	0.175	1.000	0.465	0.852	0.289	0.366	0.465	0.165	0.465	0.028	0.326	0.000	0.048	0.336	0.306	0.289	0.255	0.465	0.322	0.465	0.329	0.465	0.016	0.370	0.000	0.000
InternetService_No	0.227	0.036	0.218	0.139	1.000	0.380	0.465	1.000	0.967	0.171	0.210	1.000	0.381	1.000	0.333	0.320	0.000	0.000	0.284	0.321	0.171	0.182	1.000	0.418	1.000	0.415	1.000	0.336	0.429	0.000	0.023
MonthlyCharges	0.276	0.128	0.303	0.142	0.967	0.519	0.852	0.967	1.000	0.664	0.526	0.967	0.488	0.967	0.425	0.360	0.156	0.056	0.318	0.377	0.664	0.234	0.967	0.665	0.967	0.668	0.967	0.445	0.638	0.007	0.276
MultipleLines_No phone service	0.000	0.000	0.000	0.000	0.171	0.070	0.289	0.171	0.664	1.000	0.279	0.171	0.050	0.171	0.092	0.011	0.012	0.000	0.000	0.000	0.999	0.000	0.171	0.030	0.171	0.019	0.171	0.095	0.152	0.000	0.000
MultipleLines_Yes	0.038	0.000	0.105	0.021	0.210	0.200	0.366	0.210	0.526	0.279	1.000	0.210	0.202	0.210	0.097	0.163	0.141	0.059	0.082	0.227	0.279	0.142	0.210	0.258	0.210	0.257	0.210	0.100	0.469	0.000	0.333
OnlineBackup_No internet service	0.227	0.036	0.218	0.139	1.000	0.380	0.465	1.000	0.967	0.171	0.210	1.000	0.381	1.000	0.333	0.320	0.000	0.000	0.284	0.321	0.171	0.182	1.000	0.418	1.000	0.415	1.000	0.336	0.429	0.000	0.023
OnlineBackup_Yes	0.081	0.083	0.110	0.020	0.381	0.303	0.165	0.381	0.488	0.050	0.202	0.381	1.000	0.381	0.283	0.126	0.141	0.090	0.000	0.173	0.050	0.065	0.381	0.274	0.381	0.282	0.381	0.294	0.508	0.006	0.358
OnlineSecurity_No internet service	0.227	0.036	0.218	0.139	1.000	0.380	0.465	1.000	0.967	0.171	0.210	1.000	0.381	1.000	0.333	0.320	0.000	0.000	0.284	0.321	0.171	0.182	1.000	0.418	1.000	0.415	1.000	0.336	0.429	0.000	0.023
OnlineSecurity_Yes	0.170	0.099	0.191	0.080	0.333	0.275	0.028	0.333	0.425	0.092	0.097	0.333	0.283	0.333	1.000	0.000	0.142	0.115	0.111	0.080	0.092	0.036	0.333	0.187	0.333	0.175	0.333	0.354	0.420	0.012	0.326
PaperlessBilling	0.191	0.050	0.147	0.110	0.320	0.103	0.326	0.320	0.360	0.011	0.163	0.320	0.126	0.320	0.000	1.000	0.008	0.006	0.208	0.205	0.011	0.156	0.320	0.211	0.320	0.223	0.320	0.036	0.159	0.000	0.000
Partner	0.150	0.082	0.247	0.452	0.000	0.153	0.000	0.000	0.156	0.012	0.141	0.000	0.141	0.000	0.142	0.008	1.000	0.081	0.083	0.094	0.012	0.011	0.000	0.117	0.000	0.124	0.000	0.119	0.326	0.000	0.378
PaymentMethod_Credit card (automatic)	0.133	0.066	0.172	0.059	0.000	0.111	0.048	0.000	0.056	0.000	0.059	0.000	0.090	0.000	0.115	0.006	0.081	1.000	0.373	0.285	0.000	0.020	0.000	0.047	0.000	0.038	0.000	0.116	0.193	0.000	0.234
PaymentMethod_Electronic check	0.301	0.108	0.282	0.150	0.284	0.000	0.336	0.284	0.318	0.000	0.082	0.284	0.000	0.284	0.111	0.208	0.083	0.373	1.000	0.387	0.000	0.171	0.284	0.137	0.284	0.144	0.284	0.114	0.080	0.000	0.212
PaymentMethod_Mailed check	0.091	0.000	0.000	0.057	0.321	0.187	0.306	0.321	0.377	0.000	0.227	0.321	0.173	0.321	0.080	0.205	0.094	0.285	0.387	1.000	0.000	0.153	0.321	0.250	0.321	0.247	0.321	0.084	0.315	0.006	0.235
PhoneService	0.000	0.000	0.000	0.000	0.171	0.070	0.289	0.171	0.664	0.999	0.279	0.171	0.050	0.171	0.092	0.011	0.012	0.000	0.000	0.000	1.000	0.000	0.171	0.030	0.171	0.019	0.171	0.095	0.152	0.000	0.000
SeniorCitizen	0.150	0.044	0.116	0.210	0.182	0.058	0.255	0.182	0.234	0.000	0.142	0.182	0.065	0.182	0.036	0.156	0.011	0.020	0.171	0.153	0.000	1.000	0.182	0.119	0.182	0.104	0.182	0.059	0.114	0.000	0.022
StreamingMovies_No internet service	0.227	0.036	0.218	0.139	1.000	0.380	0.465	1.000	0.967	0.171	0.210	1.000	0.381	1.000	0.333	0.320	0.000	0.000	0.284	0.321	0.171	0.182	1.000	0.418	1.000	0.415	1.000	0.336	0.429	0.000	0.023
StreamingMovies_Yes	0.060	0.063	0.073	0.038	0.418	0.402	0.322	0.418	0.665	0.030	0.258	0.418	0.274	0.418	0.187	0.211	0.117	0.047	0.137	0.250	0.030	0.119	0.418	1.000	0.418	0.533	0.418	0.279	0.518	0.000	0.285
StreamingTV_No internet service	0.227	0.036	0.218	0.139	1.000	0.380	0.465	1.000	0.967	0.171	0.210	1.000	0.381	1.000	0.333	0.320	0.000	0.000	0.284	0.321	0.171	0.182	1.000	0.418	1.000	0.415	1.000	0.336	0.429	0.000	0.023
StreamingTV_Yes	0.062	0.060	0.071	0.011	0.415	0.390	0.329	0.415	0.668	0.019	0.257	0.415	0.282	0.415	0.175	0.223	0.124	0.038	0.144	0.247	0.019	0.104	0.415	0.533	0.415	1.000	0.415	0.278	0.513	0.000	0.276
TechSupport_No internet service	0.227	0.036	0.218	0.139	1.000	0.380	0.465	1.000	0.967	0.171	0.210	1.000	0.381	1.000	0.333	0.320	0.000	0.000	0.284	0.321	0.171	0.182	1.000	0.418	1.000	0.415	1.000	0.336	0.429	0.000	0.023
TechSupport_Yes	0.164	0.095	0.240	0.062	0.336	0.333	0.016	0.336	0.445	0.095	0.100	0.336	0.294	0.336	0.354	0.036	0.119	0.116	0.114	0.084	0.095	0.059	0.336	0.279	0.336	0.278	0.336	1.000	0.437	0.000	0.324
TotalCharges	0.213	0.208	0.396	0.091	0.429	0.522	0.370	0.429	0.638	0.152	0.469	0.429	0.508	0.429	0.420	0.159	0.326	0.193	0.080	0.315	0.152	0.114	0.429	0.518	0.429	0.513	0.429	0.437	1.000	0.000	0.890
gender	0.000	0.000	0.000	0.000	0.000	0.000	0.000	0.000	0.007	0.000	0.000	0.000	0.006	0.000	0.012	0.000	0.000	0.000	0.000	0.006	0.000	0.000	0.000	0.000	0.000	0.000	0.000	0.000	0.000	1.000	0.000
tenure	0.364	0.307	0.596	0.163	0.023	0.358	0.000	0.023	0.276	0.000	0.333	0.023	0.358	0.023	0.326	0.000	0.378	0.234	0.212	0.235	0.000	0.022	0.023	0.285	0.023	0.276	0.023	0.324	0.890	0.000	1.000

	gender	Partner	Dependents	tenure	PhoneService	PaperlessBilling	MonthlyCharges	TotalCharges	Churn	InternetService_Fiber optic	InternetService_No	Contract_One year	Contract_Two year	PaymentMethod_Credit card (automatic)	PaymentMethod_Electronic check	PaymentMethod_Mailed check	MultipleLines_No phone service	MultipleLines_Yes	OnlineSecurity_No internet service	OnlineSecurity_Yes	OnlineBackup_No internet service	OnlineBackup_Yes	DeviceProtection_No internet service	DeviceProtection_Yes	TechSupport_No internet service	TechSupport_Yes	StreamingTV_No internet service	StreamingTV_Yes	StreamingMovies_No internet service	StreamingMovies_Yes
0	0	1	0	1	0	1	29.85	29.85	0	False	False	False	False	False	True	False	True	False	False	False	False	True	False	False	False	False	False	False	False	False
1	1	0	0	34	1	0	56.95	1889.50	0	False	False	True	False	False	False	True	False	False	False	True	False	False	False	True	False	False	False	False	False	False
2	1	0	0	2	1	1	53.85	108.15	1	False	False	False	False	False	False	True	False	False	False	True	False	True	False	False	False	False	False	False	False	False
3	1	0	0	45	0	0	42.30	1840.75	0	False	False	True	False	False	False	False	True	False	False	True	False	False	False	True	False	True	False	False	False	False
4	0	0	0	2	1	1	70.70	151.65	1	True	False	False	False	False	True	False	False	False	False	False	False	False	False	False	False	False	False	False	False	False
5	0	0	0	8	1	1	99.65	820.50	1	True	False	False	False	False	True	False	False	True	False	False	False	False	False	True	False	False	False	True	False	True
6	1	0	1	22	1	1	89.10	1949.40	0	True	False	False	False	True	False	False	False	True	False	False	False	True	False	False	False	False	False	True	False	False
7	0	0	0	10	0	0	29.75	301.90	0	False	False	False	False	False	False	True	True	False	False	True	False	False	False	False	False	False	False	False	False	False
8	0	1	0	28	1	1	104.80	3046.05	1	True	False	False	False	False	True	False	False	True	False	False	False	False	False	True	False	True	False	True	False	True
9	1	0	1	62	1	0	56.15	3487.95	0	False	False	True	False	False	False	False	False	False	False	True	False	True	False	False	False	False	False	False	False	False

	gender	SeniorCitizen	Partner	Dependents	tenure	PhoneService	PaperlessBilling	MonthlyCharges	TotalCharges	Churn	InternetService_Fiber optic	InternetService_No	Contract_One year	Contract_Two year	PaymentMethod_Credit card (automatic)	PaymentMethod_Electronic check	PaymentMethod_Mailed check	MultipleLines_No phone service	MultipleLines_Yes	OnlineSecurity_No internet service	OnlineSecurity_Yes	OnlineBackup_No internet service	OnlineBackup_Yes	DeviceProtection_No internet service	DeviceProtection_Yes	TechSupport_No internet service	TechSupport_Yes	StreamingTV_No internet service	StreamingTV_Yes	StreamingMovies_No internet service	StreamingMovies_Yes
7033	1	0	0	0	38	1	1	69.50	2625.25	0	True	False	False	False	True	False	False	False	False	False	False	False	False	False	False	False	False	False	False	False	False
7034	0	0	0	0	67	1	1	102.95	6886.25	1	True	False	False	False	True	False	False	False	True	False	True	False	True	False	True	False	False	False	True	False	False
7035	1	0	0	0	19	1	1	78.70	1495.10	0	True	False	False	False	False	False	False	False	False	False	False	False	False	False	False	False	False	False	True	False	False
7036	0	0	0	0	12	0	0	60.65	743.30	0	False	False	True	False	False	True	False	True	False	False	False	False	True	False	True	False	True	False	True	False	True
7037	0	0	0	0	72	1	1	21.15	1419.40	0	False	True	False	True	False	False	False	False	False	True	False	True	False	True	False	True	False	True	False	True	False
7038	1	0	1	1	24	1	1	84.80	1990.50	0	False	False	True	False	False	False	True	False	True	False	True	False	False	False	True	False	True	False	True	False	True
7039	0	0	1	1	72	1	1	103.20	7362.90	0	True	False	True	False	True	False	False	False	True	False	False	False	True	False	True	False	False	False	True	False	True
7040	0	0	1	1	11	0	1	29.60	346.45	0	False	False	False	False	False	True	False	True	False	False	True	False	False	False	False	False	False	False	False	False	False
7041	1	1	1	0	4	1	1	74.40	306.60	1	True	False	False	False	False	False	True	False	True	False	False	False	False	False	False	False	False	False	False	False	False
7042	1	0	0	0	66	1	1	105.65	6844.50	0	True	False	False	True	False	False	False	False	False	False	True	False	False	False	True	False	True	False	True	False	True

	gender	SeniorCitizen	tenure	PhoneService	PaperlessBilling	MonthlyCharges	TotalCharges	Churn	InternetService_Fiber optic	InternetService_No	Contract_One year	Contract_Two year	PaymentMethod_Credit card (automatic)	PaymentMethod_Electronic check	PaymentMethod_Mailed check	MultipleLines_No phone service	MultipleLines_Yes	OnlineSecurity_No internet service	OnlineSecurity_Yes	OnlineBackup_No internet service	OnlineBackup_Yes	DeviceProtection_No internet service	DeviceProtection_Yes	TechSupport_No internet service	TechSupport_Yes	StreamingTV_No internet service	StreamingTV_Yes	StreamingMovies_No internet service	StreamingMovies_Yes	# duplicates
8	1	0	1	1	0	20.05	20.05	0	False	True	False	False	False	False	True	False	False	True	False	True	False	True	False	True	False	True	False	True	False	3
10	1	0	1	1	0	20.20	20.20	0	False	True	False	False	False	False	True	False	False	True	False	True	False	True	False	True	False	True	False	True	False	3
0	0	0	1	1	0	19.55	19.55	0	False	True	False	False	False	False	True	False	False	True	False	True	False	True	False	True	False	True	False	True	False	2
1	0	0	1	1	0	19.65	19.65	0	False	True	False	False	False	False	True	False	False	True	False	True	False	True	False	True	False	True	False	True	False	2
2	0	0	1	1	0	19.90	19.90	0	False	True	False	False	False	False	True	False	False	True	False	True	False	True	False	True	False	True	False	True	False	2
3	0	0	1	1	0	20.90	20.90	1	False	True	False	False	False	False	True	False	False	True	False	True	False	True	False	True	False	True	False	True	False	2
4	0	0	1	1	1	69.20	69.20	1	True	False	False	False	False	True	False	False	False	False	False	False	False	False	False	False	False	False	False	False	False	2
5	0	0	1	1	1	70.10	70.10	1	True	False	False	False	False	True	False	False	False	False	False	False	False	False	False	False	False	False	False	False	False	2
6	0	0	1	1	1	70.15	70.15	1	True	False	False	False	False	False	True	False	False	False	False	False	False	False	False	False	False	False	False	False	False	2
7	0	1	1	1	1	69.60	69.60	1	True	False	False	False	False	True	False	False	False	False	False	False	False	False	False	False	False	False	False	False	False	2

Number of variables	21
Number of observations	7032
Missing cells	0
Missing cells (%)	0.0%
Total size in memory	1.1 MiB
Average record size in memory	170.2 B

`customerID` has a high cardinality: 7032 distinct values	High cardinality
`PhoneService` is highly imbalanced (54.2%)	Imbalance
`customerID` has unique values	Unique
`SeniorCitizen` has 5890 (83.8%) zeros	Zeros
`Churn` has 5163 (73.4%) zeros	Zeros

Analysis started	2025-08-05 04:38:52.759254
Analysis finished	2025-08-05 04:38:52.870863
Duration	0.11 seconds
Software version	ydata-profiling vv4.16.1
Download configuration	config.json

Telco完整项目展示

Tableau可视化展示

PPT报告展示

Telco业分析报告

Overview

Variables

Common Values

Length

Common Values (Plot)

Most occurring characters

Most occurring categories

Most frequent character per category

(unknown)

Most occurring scripts

Most frequent character per script

(unknown)

Most occurring blocks

Most frequent character per block

(unknown)

Common Values

Length

Common Values (Plot)

Most occurring characters

Most occurring categories

Most frequent character per category

(unknown)

Most occurring scripts

Most frequent character per script

(unknown)

Most occurring blocks

Most frequent character per block

(unknown)

Common Values

Length

Common Values (Plot)

Most occurring characters

Most occurring categories

Most frequent character per category

(unknown)

Most occurring scripts

Most frequent character per script

(unknown)

Most occurring blocks

Most frequent character per block

(unknown)

Common Values

Length

Common Values (Plot)

Most occurring characters

Most occurring categories

Most frequent character per category

(unknown)

Most occurring scripts

Most frequent character per script

(unknown)

Most occurring blocks

Most frequent character per block

(unknown)

Common Values

Length

Common Values (Plot)

Most occurring characters

Most occurring categories

Most frequent character per category

(unknown)

Most occurring scripts

Most frequent character per script

(unknown)

Most occurring blocks

Most frequent character per block

(unknown)

Common Values

Length

Common Values (Plot)

Most occurring characters

Most occurring categories

Most frequent character per category

(unknown)

Most occurring scripts

Most frequent character per script

0002-ORFBO	1
6619-RPLQZ	1
6629-LADHQ	1
6629-CZTTH	1
6627-CFOSN	1
Other values (7027)	7027

1st row	7590-VHVEG
2nd row	5575-GNVDE
3rd row	3668-QPYBK
4th row	7795-CFOCW
5th row	9237-HQITU

Standard deviation	0.3688439968
Coefficient of variation (CV)	2.271200512
Kurtosis	1.353321348
Mean	0.1624004551
Median Absolute Deviation (MAD)	0
Skewness	1.831102544
Sum	1142
Variance	0.1360458939
Monotonicity	Not monotonic

Standard deviation	24.54525971
Coefficient of variation (CV)	0.7570606881
Kurtosis	-1.38782258
Mean	32.42178612
Median Absolute Deviation (MAD)	22
Skewness	0.2377308319
Sum	227990
Variance	602.4697742
Monotonicity	Not monotonic

1st row	No phone service
2nd row	No
3rd row	No
4th row	No phone service
5th row	No

1st row	Month-to-month
2nd row	One year
3rd row	Month-to-month
4th row	One year
5th row	Month-to-month

Distinct	4
Distinct (%)	0.1%
Missing	0
Missing (%)	0.0%
Memory size	62.3 KiB

Electronic check	2365
Mailed check	1604
Bank transfer (automatic)	1542
Credit card (automatic)	1521

Distinct	1584
Distinct (%)	22.5%
Missing	0
Missing (%)	0.0%
Infinite	0
Infinite (%)	0.0%
Mean	64.79820819

Standard deviation	30.08597388
Coefficient of variation (CV)	0.464302559
Kurtosis	-1.256156424
Mean	64.79820819
Median Absolute Deviation (MAD)	24.05
Skewness	-0.2221029277
Sum	455661
Variance	905.1658246
Monotonicity	Not monotonic